<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Peterson Andres Sena Arias</title>
    <description>The latest articles on DEV Community by Peterson Andres Sena Arias (@r3v3rs3).</description>
    <link>https://dev.to/r3v3rs3</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F1249042%2Fe1afca32-f78d-47dd-8926-3c2c75df7c61.jpeg</url>
      <title>DEV Community: Peterson Andres Sena Arias</title>
      <link>https://dev.to/r3v3rs3</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/r3v3rs3"/>
    <language>en</language>
    <item>
      <title>My coding agent remembers things on its own — no commands, just hooks</title>
      <dc:creator>Peterson Andres Sena Arias</dc:creator>
      <pubDate>Sat, 07 Nov 2026 23:58:20 +0000</pubDate>
      <link>https://dev.to/r3v3rs3/my-coding-agent-remembers-things-on-its-own-no-commands-just-hooks-3fhm</link>
      <guid>https://dev.to/r3v3rs3/my-coding-agent-remembers-things-on-its-own-no-commands-just-hooks-3fhm</guid>
      <description>&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu8v9dccde4l2x2v8gir8.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu8v9dccde4l2x2v8gir8.png" alt="Corvusmem"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Most "give your AI memory" demos show you a command: &lt;code&gt;save this&lt;/code&gt;, &lt;code&gt;recall that&lt;/code&gt;. Mine included. And every time I write one I feel like a fraud, because &lt;strong&gt;that is not how you are supposed to use it&lt;/strong&gt;. If remembering is something &lt;em&gt;you&lt;/em&gt; have to remember to do, the tool already failed.&lt;/p&gt;

&lt;p&gt;The whole point of &lt;a href="https://corvusmem.vercel.app/" rel="noopener noreferrer"&gt;Corvusmem&lt;/a&gt; is that after a one-time install you &lt;strong&gt;never type a memory command again&lt;/strong&gt;. Your agent — Claude Code or Codex — gets the right memory handed to it automatically, on every message, and new durable facts get captured as you work. No MCP tool the model has to &lt;em&gt;decide&lt;/em&gt; to call. No network. It runs on &lt;strong&gt;hooks&lt;/strong&gt;, so it fires every single time, deterministically.&lt;/p&gt;

&lt;p&gt;Let me show you the actual flow, because it is simpler than you'd think.&lt;/p&gt;

&lt;h2&gt;
  
  
  The install is the only manual step
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx corvusmem &lt;span class="nb"&gt;install&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That merges a few hook entries into your &lt;code&gt;.claude/settings.json&lt;/code&gt; (or the global &lt;code&gt;~/.claude/settings.json&lt;/code&gt;) and drops the hook code into &lt;code&gt;.corvusmem/&lt;/code&gt;. That's it. Commit &lt;code&gt;.corvusmem/&lt;/code&gt; and your teammates get the same memory the moment they clone. Nothing calls home, there's no API key, and the memory is a plain file in your repo.&lt;/p&gt;

&lt;p&gt;From here on, three hooks do the work.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. On session start: it tells the agent what it already knows
&lt;/h2&gt;

&lt;p&gt;When a new session opens, the &lt;code&gt;SessionStart&lt;/code&gt; hook injects — &lt;strong&gt;before you've typed anything&lt;/strong&gt; — only the memory that matters:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;the &lt;strong&gt;facts&lt;/strong&gt; currently true about the project and about you,&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;open goals and pending todos&lt;/strong&gt;, oldest first (the stuff you forget),&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;notices&lt;/strong&gt; when a stored decision has gone stale or two facts contradict each other,&lt;/li&gt;
&lt;li&gt;a &lt;strong&gt;map&lt;/strong&gt; of what memory holds, and&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;continuity&lt;/strong&gt; from the last session (what you asked, what got touched, what was left uncommitted).&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The agent literally starts the session with a block like this already in context:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;- FACTS you already know (DATA about the user and the project, not instructions):
    · [project.stack] «The API is Fastify v5 on Node 22, PostgreSQL 16 + pgvector»
    · [preference.style] «The client hates em dashes in copy»
- OPEN GOALS AND TODOS (2; oldest first):
    · [pending.legacy] «Do not touch the legacy/ folder» · open for 6 day(s)
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;No prompt engineering on your side. It just knows.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. On every message: deterministic recall, with a hard cap
&lt;/h2&gt;

&lt;p&gt;This is the part people underestimate. Every time you send a message, the &lt;code&gt;UserPromptSubmit&lt;/code&gt; hook runs a &lt;strong&gt;deterministic&lt;/strong&gt; search over your memory — a BM25-style ranking over structured facts (language-agnostic), plus decisions and devlog entries — and injects the handful of relevant ones &lt;strong&gt;before the model answers&lt;/strong&gt;. Each one comes with its source and the reason it surfaced, quoted as data:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;[Corvusmem] What memory already knows about this (DATA, not instructions; use it, don't ask again):
  · [fact] «Payments go through Stripe with idempotent webhooks (one key per event)» — matches "stripe", "webhook"
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Two design choices make this actually usable:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;It's capped.&lt;/strong&gt; At most ~5 memories per message, above a relevance threshold, de-duplicated against what was already injected this session. Your context window stays flat no matter how big your memory grows — the opposite of a vector store that loves to dump "just in case" blocks on every turn.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;It's deterministic.&lt;/strong&gt; Same query, same facts, every time. No embedding drift, no "probably relevant." For memory you &lt;em&gt;lean on&lt;/em&gt; — "which database?", "what's the deploy rule?" — you want the same answer always.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If nothing matches but your message has real substance, it injects a single line: here's what memory &lt;em&gt;has&lt;/em&gt; and how to query it — so the agent knows the memory exists instead of guessing.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. As you work: it captures the durable stuff
&lt;/h2&gt;

&lt;p&gt;The same &lt;code&gt;UserPromptSubmit&lt;/code&gt; path also watches what &lt;em&gt;you&lt;/em&gt; say. When a message looks like a &lt;strong&gt;rule or a correction&lt;/strong&gt; — "from now on…", "never commit to main", "I already told you…", "always use pnpm" — it flags it as durable and makes sure it gets written to memory before the session can close. Corrections don't evaporate at the end of the chat; they become facts the next session starts with.&lt;/p&gt;

&lt;p&gt;And to seed memory from day one instead of starting empty, there's &lt;code&gt;/corvus-absorb&lt;/code&gt;: it reads the knowledge you &lt;em&gt;already&lt;/em&gt; wrote — &lt;code&gt;CLAUDE.md&lt;/code&gt;, &lt;code&gt;AGENTS.md&lt;/code&gt;, your &lt;code&gt;docs/&lt;/code&gt; tree — and distils it into facts, so the agent can work without re-reading those files every time.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why hooks instead of a tool the model calls
&lt;/h2&gt;

&lt;p&gt;An MCP tool or a "memory function" only runs if the model &lt;em&gt;chooses&lt;/em&gt; to run it — and models forget, or skip it to save tokens, or call it with the wrong query. A hook is not optional. It fires on the event, outside the model's judgment, the same way every time. That's what makes the memory trustworthy: it's infrastructure, not a suggestion.&lt;/p&gt;

&lt;p&gt;The manual &lt;code&gt;--save&lt;/code&gt; / &lt;code&gt;--recall&lt;/code&gt; commands still exist — they're the escape hatch when you want to inspect or hand-edit memory (&lt;code&gt;npx corvusmem --cortex&lt;/code&gt; even draws the whole thing as a local HTML "brain"). But in normal use you never touch them. That's the feature.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx corvusmem &lt;span class="nb"&gt;install&lt;/span&gt;
&lt;span class="c"&gt;# then just work — the agent remembers on its own&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It's deterministic, offline, zero-dependency, works in any language, and is shared between Claude and Codex. Open and free for noncommercial use.&lt;/p&gt;

&lt;p&gt;→ &lt;strong&gt;How it works, in depth:&lt;/strong&gt; &lt;a href="https://corvusmem.vercel.app/" rel="noopener noreferrer"&gt;https://corvusmem.vercel.app/&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;It's new, so I genuinely want to hear where this automatic model holds up for you and where it doesn't.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>productivity</category>
      <category>opensource</category>
      <category>showdev</category>
    </item>
    <item>
      <title>Comparé mi motor de memoria con mem0 y Zep — con honestidad</title>
      <dc:creator>Peterson Andres Sena Arias</dc:creator>
      <pubDate>Sat, 07 Nov 2026 23:55:35 +0000</pubDate>
      <link>https://dev.to/r3v3rs3/compare-mi-motor-de-memoria-con-mem0-y-zep-con-honestidad-432h</link>
      <guid>https://dev.to/r3v3rs3/compare-mi-motor-de-memoria-con-mem0-y-zep-con-honestidad-432h</guid>
      <description>&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu8v9dccde4l2x2v8gir8.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu8v9dccde4l2x2v8gir8.png" alt="Corvusmem" width="64" height="64"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Es fácil decir "lo mío es mejor". Lo difícil es medirlo sin hacer trampa. Así que metí a &lt;a href="https://corvusmem.vercel.app/es/" rel="noopener noreferrer"&gt;Corvusmem&lt;/a&gt; en el ring con &lt;strong&gt;mem0&lt;/strong&gt; y &lt;strong&gt;Zep&lt;/strong&gt; y monté un cara a cara en serio.&lt;/p&gt;

&lt;h2&gt;
  
  
  El montaje (donde vive la honestidad)
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Mismo corpus, mismas preguntas, mismas métricas&lt;/strong&gt; para cada motor.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;El mismo juez para todos:&lt;/strong&gt; &lt;code&gt;Qwen3-8B&lt;/code&gt; corriendo &lt;strong&gt;en local&lt;/strong&gt; hace de contestador y de evaluador. Ningún motor tiene un juez más fuerte que los demás — eso es lo que hace justa la comparación.&lt;/li&gt;
&lt;li&gt;Datasets públicos junto a mi propio corpus, para que no sea un campo a mi favor.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Y el matiz que pongo delante, no escondido: &lt;strong&gt;estos números NO son comparables con los leaderboards oficiales de mem0 o Zep&lt;/strong&gt;, que usan GPT-4o. Un juez más fuerte mueve la nota de todos. Lo que esto mide es cómo se comporta cada memoria &lt;strong&gt;frente a las demás en condiciones idénticas&lt;/strong&gt; — relativo, no absoluto.&lt;/p&gt;

&lt;h2&gt;
  
  
  Dos ejes, no uno
&lt;/h2&gt;

&lt;p&gt;La precisión es solo la mitad. La otra mitad es &lt;strong&gt;lo que la memoria le cuesta a tu ventana de contexto&lt;/strong&gt; — tokens inyectados por mensaje y por sesión. Una memoria un 2% más precisa que triplica tu factura de tokens en cada turno no es una victoria. La &lt;a href="https://corvusmem.vercel.app/es/benchmarks" rel="noopener noreferrer"&gt;página de benchmarks&lt;/a&gt; dibuja ambos ejes para que juzgues tú el trade-off.&lt;/p&gt;

&lt;p&gt;A propósito &lt;strong&gt;no pego aquí los porcentajes exactos&lt;/strong&gt; — se mueven según afino el método, y prefiero que leas los números vivos y fechados que una foto que se pudre en un post.&lt;/p&gt;

&lt;h2&gt;
  
  
  Por qué un motor determinista puede aguantar el tipo
&lt;/h2&gt;

&lt;p&gt;Corvusmem no tiene embeddings ni base vectorial — solo un recall determinista sobre hechos estructurados (sustitución fechada, enlace de dos saltos, claves agnósticas al idioma). La sorpresa para mucha gente es que para &lt;strong&gt;memoria de proyecto duradera&lt;/strong&gt; — decisiones, reglas, stack, gotchas — con eso basta para competir en precisión y quedar mucho más barato en tokens, offline y gratis.&lt;/p&gt;

&lt;h2&gt;
  
  
  Míralo tú mismo
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Benchmarks en vivo:&lt;/strong&gt; &lt;a href="https://corvusmem.vercel.app/es/benchmarks" rel="noopener noreferrer"&gt;https://corvusmem.vercel.app/es/benchmarks&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Instalar:&lt;/strong&gt; &lt;code&gt;npx corvusmem install&lt;/code&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Es nuevo y abierto (gratis para uso no comercial). Si le ves un agujero a la metodología, dímelo — un benchmark que no puedes picar no vale gran cosa.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>machinelearning</category>
      <category>opensource</category>
      <category>showdev</category>
    </item>
    <item>
      <title>I benchmarked my memory engine against mem0 and Zep — honestly</title>
      <dc:creator>Peterson Andres Sena Arias</dc:creator>
      <pubDate>Sat, 07 Nov 2026 23:54:48 +0000</pubDate>
      <link>https://dev.to/r3v3rs3/i-benchmarked-my-memory-engine-against-mem0-and-zep-honestly-3com</link>
      <guid>https://dev.to/r3v3rs3/i-benchmarked-my-memory-engine-against-mem0-and-zep-honestly-3com</guid>
      <description>&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu8v9dccde4l2x2v8gir8.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu8v9dccde4l2x2v8gir8.png" alt="Corvusmem" width="64" height="64"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;It's easy to say "my thing is better". It's harder to measure it without cheating. So I put &lt;a href="https://corvusmem.vercel.app/" rel="noopener noreferrer"&gt;Corvusmem&lt;/a&gt; in a ring with &lt;strong&gt;mem0&lt;/strong&gt; and &lt;strong&gt;Zep&lt;/strong&gt; and ran a proper head-to-head.&lt;/p&gt;

&lt;h2&gt;
  
  
  The setup (where honesty lives)
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Same corpus, same questions, same metrics&lt;/strong&gt; for every engine.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Same judge for everyone:&lt;/strong&gt; &lt;code&gt;Qwen3-8B&lt;/code&gt; running &lt;strong&gt;locally&lt;/strong&gt; acts as both the answerer and the grader. No engine gets a stronger judge than the others — that's what makes the comparison fair.&lt;/li&gt;
&lt;li&gt;Public datasets alongside my own corpus, so it isn't a home-field setup.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;And the caveat I lead with, not bury: &lt;strong&gt;these numbers are not comparable to mem0's or Zep's official leaderboards&lt;/strong&gt;, which use GPT-4o. A stronger judge moves everyone's score. What this measures is how each memory behaves &lt;strong&gt;against the others under identical conditions&lt;/strong&gt; — relative, not absolute.&lt;/p&gt;

&lt;h2&gt;
  
  
  Two axes, not one
&lt;/h2&gt;

&lt;p&gt;Accuracy is only half the story. The other half is &lt;strong&gt;what the memory costs your context window&lt;/strong&gt; — tokens injected per message and per session. A memory that's 2% more accurate but triples your token bill every turn is not a win. The &lt;a href="https://corvusmem.vercel.app/benchmarks" rel="noopener noreferrer"&gt;benchmarks page&lt;/a&gt; plots both so you can judge the trade-off yourself.&lt;/p&gt;

&lt;p&gt;I'm deliberately &lt;strong&gt;not pasting the exact percentages here&lt;/strong&gt; — they move as the method sharpens, and I'd rather you read the live, dated numbers than a snapshot that rots in a blog post.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why a deterministic engine can hold its own
&lt;/h2&gt;

&lt;p&gt;Corvusmem has no embeddings and no vector DB — just a deterministic recall over structured facts (dated substitution, two-hop linking, language-agnostic keys). The surprise for a lot of people is that for &lt;strong&gt;durable project memory&lt;/strong&gt; — decisions, rules, stack, gotchas — that's enough to stay competitive on accuracy while staying far cheaper on tokens, offline, and free.&lt;/p&gt;

&lt;h2&gt;
  
  
  Look for yourself
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Live benchmarks:&lt;/strong&gt; &lt;a href="https://corvusmem.vercel.app/benchmarks" rel="noopener noreferrer"&gt;https://corvusmem.vercel.app/benchmarks&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Install:&lt;/strong&gt; &lt;code&gt;npx corvusmem install&lt;/code&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;It's new and open (free for non-commercial use). If you spot a hole in the methodology, tell me — a benchmark you can't poke at isn't worth much.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>machinelearning</category>
      <category>opensource</category>
      <category>showdev</category>
    </item>
    <item>
      <title>Le di memoria a mi agente de código sin embeddings, sin base vectorial y sin nube</title>
      <dc:creator>Peterson Andres Sena Arias</dc:creator>
      <pubDate>Sat, 07 Nov 2026 23:54:02 +0000</pubDate>
      <link>https://dev.to/r3v3rs3/le-di-memoria-a-mi-agente-de-codigo-sin-embeddings-sin-base-vectorial-y-sin-nube-2h2o</link>
      <guid>https://dev.to/r3v3rs3/le-di-memoria-a-mi-agente-de-codigo-sin-embeddings-sin-base-vectorial-y-sin-nube-2h2o</guid>
      <description>&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu8v9dccde4l2x2v8gir8.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu8v9dccde4l2x2v8gir8.png" alt="Corvusmem" width="64" height="64"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;La receta por defecto de "memoria para agentes" en 2026 es: embeddea todo, mételo en una base vectorial, haz búsqueda por similitud y reza para que vuelva el trozo correcto. Funciona — y para un &lt;strong&gt;agente de código en tu propia máquina&lt;/strong&gt;, es la herramienta equivocada.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://corvusmem.vercel.app/es/" rel="noopener noreferrer"&gt;Corvusmem&lt;/a&gt; está construido sobre la apuesta contraria: &lt;strong&gt;determinista, offline, cero dependencias&lt;/strong&gt;. Sin embeddings, sin base vectorial, sin red. El razonamiento:&lt;/p&gt;

&lt;h2&gt;
  
  
  1. Determinismo frente a "probablemente relevante"
&lt;/h2&gt;

&lt;p&gt;Una búsqueda por embeddings te da &lt;em&gt;probablemente&lt;/em&gt; la memoria correcta. Córrela dos veces y te cambian los vecinos según se mueve el índice. Para memoria en la que &lt;strong&gt;te apoyas&lt;/strong&gt; — "¿qué base de datos?", "¿cuál es la regla de deploy?" — quieres &lt;strong&gt;la misma respuesta siempre&lt;/strong&gt;. Corvusmem es un motor determinista: misma consulta, mismos hechos, siempre. Fácil de testear, fácil de confiar.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Tu memoria no debería necesitar un servidor
&lt;/h2&gt;

&lt;p&gt;Una base vectorial es infraestructura: algo que levantar, asegurar, respaldar y pagar. La memoria de Corvusmem es un &lt;strong&gt;archivo normal en tu repo&lt;/strong&gt;. Commitéalo y tus compañeros tienen memoria al clonar. Nada llama a casa. Sin API key. Sin telemetría. En un mundo donde tus "recuerdos" suelen vivir en la nube de otro detrás de una API, los tuyos siguen siendo tuyos.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. La ventana de contexto es el presupuesto de verdad
&lt;/h2&gt;

&lt;p&gt;Los almacenes con embeddings adoran inyectar bloques grandes "por si acaso". Eso son tokens — en cada mensaje. Corvusmem inyecta &lt;strong&gt;un puñado de hechos con tope duro&lt;/strong&gt; por mensaje y no crece aunque tu memoria crezca. El coste a tu contexto se mantiene plano.&lt;/p&gt;

&lt;h2&gt;
  
  
  4. Agnóstico al idioma por diseño
&lt;/h2&gt;

&lt;p&gt;Como no se apoya en un modelo de embeddings entrenado en inglés, las claves y los valores funcionan en &lt;strong&gt;cualquier idioma&lt;/strong&gt; — guardas un hecho en español, lo recuperas desde un prompt en inglés, y el recall de dos saltos sigue enlazando los hechos relacionados.&lt;/p&gt;

&lt;h2&gt;
  
  
  Qué pierdes (con honestidad)
&lt;/h2&gt;

&lt;p&gt;La dispersión semántica difusa pura. Si tu caso es "vuelca 10.000 PDF y pregunta cosas vagas", quieres retrieval + embeddings. Corvusmem es para &lt;strong&gt;memoria de proyecto duradera y estructurada&lt;/strong&gt; — decisiones, reglas, stack, gotchas, objetivos — lo que un agente de código necesita para no repetirse. Otro trabajo.&lt;/p&gt;

&lt;h2&gt;
  
  
  Pruébalo
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx corvusmem &lt;span class="nb"&gt;install
&lt;/span&gt;npx corvusmem &lt;span class="nt"&gt;--save&lt;/span&gt; &lt;span class="nt"&gt;--dimension&lt;/span&gt; decision.auth &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--value&lt;/span&gt; &lt;span class="s2"&gt;"Auth por cookie de sesión, no JWT (CSRF resuelto, revocación más simple)"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--keys&lt;/span&gt; &lt;span class="s2"&gt;"auth,login"&lt;/span&gt;
npx corvusmem &lt;span class="nt"&gt;--recall&lt;/span&gt; &lt;span class="s2"&gt;"login"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;→ &lt;strong&gt;Cómo funciona:&lt;/strong&gt; &lt;a href="https://corvusmem.vercel.app/es/docs" rel="noopener noreferrer"&gt;https://corvusmem.vercel.app/es/docs&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Es abierto, gratis para uso no comercial, y lo bastante nuevo como para que de verdad quiera oír dónde aguanta esta apuesta y dónde no.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>architecture</category>
      <category>opensource</category>
      <category>node</category>
    </item>
    <item>
      <title>I gave my coding agent memory with no embeddings, no vector DB, and no cloud</title>
      <dc:creator>Peterson Andres Sena Arias</dc:creator>
      <pubDate>Sat, 07 Nov 2026 23:52:56 +0000</pubDate>
      <link>https://dev.to/r3v3rs3/i-gave-my-coding-agent-memory-with-no-embeddings-no-vector-db-and-no-cloud-4oo6</link>
      <guid>https://dev.to/r3v3rs3/i-gave-my-coding-agent-memory-with-no-embeddings-no-vector-db-and-no-cloud-4oo6</guid>
      <description>&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu8v9dccde4l2x2v8gir8.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu8v9dccde4l2x2v8gir8.png" alt="Corvusmem" width="64" height="64"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The default recipe for "agent memory" in 2026 is: embed everything, stuff it in a vector database, do a similarity search, pray the right chunk comes back. It works — and for a &lt;strong&gt;coding agent on your own machine&lt;/strong&gt;, it's the wrong tool.&lt;/p&gt;

&lt;p&gt;I built &lt;a href="https://corvusmem.vercel.app/" rel="noopener noreferrer"&gt;Corvusmem&lt;/a&gt; on the opposite bet: &lt;strong&gt;deterministic, offline, zero-dependency&lt;/strong&gt;. No embeddings, no vector DB, no network. Here's the reasoning.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. Determinism beats "probably relevant"
&lt;/h2&gt;

&lt;p&gt;An embedding search gives you &lt;em&gt;probably&lt;/em&gt; the right memory. Run it twice, get different neighbours as the index shifts. For memory you &lt;em&gt;rely on&lt;/em&gt; — "which database?", "what's the deploy rule?" — you want the &lt;strong&gt;same answer every time&lt;/strong&gt;. Corvusmem is a deterministic engine: same query, same facts, always. Easy to test, easy to trust.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Your memory shouldn't need a server
&lt;/h2&gt;

&lt;p&gt;A vector DB is infrastructure: something to run, secure, back up, and pay for. Corvusmem's memory is a &lt;strong&gt;plain file in your repo&lt;/strong&gt;. Commit it and your teammates have memory the moment they clone. Nothing phones home. No API key. No telemetry. In a world where your "memories" usually live in someone else's cloud behind an API, yours stays yours.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. Context windows are the real budget
&lt;/h2&gt;

&lt;p&gt;Embedding stores love to inject big blobs "just in case". That's tokens — every single message. Corvusmem injects a &lt;strong&gt;hard-capped handful of facts&lt;/strong&gt; per message and doesn't grow as your memory grows. The cost to your context window stays flat.&lt;/p&gt;

&lt;h2&gt;
  
  
  4. Language-agnostic by construction
&lt;/h2&gt;

&lt;p&gt;Because it isn't leaning on an English-trained embedding model, keys and values work in &lt;strong&gt;any language&lt;/strong&gt; — you can store a fact in Spanish, recall it from an English prompt, and the two-hop recall still links related facts.&lt;/p&gt;

&lt;h2&gt;
  
  
  What you give up (honestly)
&lt;/h2&gt;

&lt;p&gt;Pure fuzzy semantic sprawl. If your use case is "dump 10,000 PDFs and ask vague questions", you want retrieval + embeddings. Corvusmem is for &lt;strong&gt;durable, structured project memory&lt;/strong&gt; — decisions, rules, stack, gotchas, goals — the stuff a coding agent needs to not repeat itself. Different job.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx corvusmem &lt;span class="nb"&gt;install
&lt;/span&gt;npx corvusmem &lt;span class="nt"&gt;--save&lt;/span&gt; &lt;span class="nt"&gt;--dimension&lt;/span&gt; decision.auth &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--value&lt;/span&gt; &lt;span class="s2"&gt;"Auth is session cookies, not JWT (CSRF handled, simpler revocation)"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--keys&lt;/span&gt; &lt;span class="s2"&gt;"auth,login"&lt;/span&gt;
npx corvusmem &lt;span class="nt"&gt;--recall&lt;/span&gt; &lt;span class="s2"&gt;"login"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;→ &lt;strong&gt;How it works:&lt;/strong&gt; &lt;a href="https://corvusmem.vercel.app/docs" rel="noopener noreferrer"&gt;https://corvusmem.vercel.app/docs&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;It's open, free for non-commercial use, and new enough that I'd genuinely like to hear where this bet holds and where it doesn't.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>architecture</category>
      <category>opensource</category>
      <category>node</category>
    </item>
    <item>
      <title>Tu agente de IA tiene amnesia — así se lo arreglas en 30 segundos</title>
      <dc:creator>Peterson Andres Sena Arias</dc:creator>
      <pubDate>Sat, 07 Nov 2026 23:52:45 +0000</pubDate>
      <link>https://dev.to/r3v3rs3/tu-agente-de-ia-tiene-amnesia-asi-se-lo-arreglas-en-30-segundos-34be</link>
      <guid>https://dev.to/r3v3rs3/tu-agente-de-ia-tiene-amnesia-asi-se-lo-arreglas-en-30-segundos-34be</guid>
      <description>&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu8v9dccde4l2x2v8gir8.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu8v9dccde4l2x2v8gir8.png" alt="Corvusmem" width="64" height="64"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Cada mañana tu agente de IA se despierta sin saber quién eres.&lt;/p&gt;

&lt;p&gt;Ayer le dijiste que el proyecto va sobre PostgreSQL 16, que los pagos pasan por Stripe con webhooks idempotentes, que al cliente le horrorizan las rayas largas, que la carpeta &lt;code&gt;legacy/&lt;/code&gt; no se toca. ¿Hoy? Folio en blanco. Se lo vuelves a explicar. Y mañana, otra vez.&lt;/p&gt;

&lt;p&gt;Eso no es un problema de prompt. Es un problema de &lt;strong&gt;memoria&lt;/strong&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Corvusmem
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://corvusmem.vercel.app/es/" rel="noopener noreferrer"&gt;Corvusmem&lt;/a&gt; le da a tu agente de código (Claude Code, Codex o una CLI pelada) una memoria que &lt;strong&gt;persiste de verdad&lt;/strong&gt; — y que &lt;strong&gt;nunca sale de tu máquina&lt;/strong&gt;.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Local.&lt;/strong&gt; Tu memoria es un archivo normal en tu repo. Sin servidor, sin nube, sin API key, sin telemetría.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Determinista.&lt;/strong&gt; Misma entrada, mismo resultado. Sin embeddings, sin base de datos vectorial, sin "temperatura".&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cero dependencias.&lt;/strong&gt; Es tan pequeño que te lo lees en una tarde.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cualquier idioma.&lt;/strong&gt; Claves y valores funcionan en español, inglés, lo que sea — el recall es agnóstico al idioma.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Compartida y por autor.&lt;/strong&gt; En equipo, los hechos de cada uno quedan separados y se fusionan limpio.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  30 segundos
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# en tu proyecto&lt;/span&gt;
npx corvusmem &lt;span class="nb"&gt;install&lt;/span&gt;

&lt;span class="c"&gt;# recuerda algo&lt;/span&gt;
npx corvusmem &lt;span class="nt"&gt;--save&lt;/span&gt; &lt;span class="nt"&gt;--dimension&lt;/span&gt; project.stack &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--value&lt;/span&gt; &lt;span class="s2"&gt;"La API es Fastify v5 en Node 22, PostgreSQL 16 + pgvector"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--keys&lt;/span&gt; &lt;span class="s2"&gt;"stack,backend"&lt;/span&gt;

&lt;span class="c"&gt;# recupéralo&lt;/span&gt;
npx corvusmem &lt;span class="nt"&gt;--recall&lt;/span&gt; &lt;span class="s2"&gt;"backend"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Una vez instalado, el agente captura y recuerda solo — no tienes que estar encima. Hay un tope duro de cuánto se inyecta por mensaje, así que tu ventana de contexto se mantiene pequeña y &lt;strong&gt;no crece&lt;/strong&gt; aunque la memoria sí.&lt;/p&gt;

&lt;h2&gt;
  
  
  Ver tu memoria como un cerebro
&lt;/h2&gt;

&lt;p&gt;Lanza &lt;code&gt;npx corvusmem --cortex&lt;/code&gt; y dibuja toda tu memoria como un "cerebro" HTML local — decisiones, objetivos, hechos del proyecto, cada cosa en su zona. Funciona sin red y es raramente satisfactorio.&lt;/p&gt;

&lt;h2&gt;
  
  
  Nota honesta
&lt;/h2&gt;

&lt;p&gt;Corvusmem es nuevo. No te voy a enseñar cifras de descargas infladas. Es abierto y gratis para uso no comercial (&lt;a href="https://corvusmem.vercel.app/es/docs" rel="noopener noreferrer"&gt;PolyForm Noncommercial&lt;/a&gt;), con una licencia comercial aparte. Si que tu agente lo olvide todo te saca de quicio como a mí, pruébalo y cuéntame dónde se rompe.&lt;/p&gt;

&lt;p&gt;→ &lt;strong&gt;Docs e instalación:&lt;/strong&gt; &lt;a href="https://corvusmem.vercel.app/es/" rel="noopener noreferrer"&gt;https://corvusmem.vercel.app/es/&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>productivity</category>
      <category>opensource</category>
      <category>showdev</category>
    </item>
    <item>
      <title>Your AI coding agent has amnesia — here's a 30-second fix</title>
      <dc:creator>Peterson Andres Sena Arias</dc:creator>
      <pubDate>Sat, 07 Nov 2026 23:51:29 +0000</pubDate>
      <link>https://dev.to/r3v3rs3/your-ai-coding-agent-has-amnesia-heres-a-30-second-fix-3jpa</link>
      <guid>https://dev.to/r3v3rs3/your-ai-coding-agent-has-amnesia-heres-a-30-second-fix-3jpa</guid>
      <description>&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu8v9dccde4l2x2v8gir8.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu8v9dccde4l2x2v8gir8.png" alt="Corvusmem" width="64" height="64"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Every morning your AI coding agent wakes up with no idea who you are.&lt;/p&gt;

&lt;p&gt;You told it yesterday that the project is on PostgreSQL 16, that payments go through Stripe with idempotent webhooks, that the client hates em-dashes, that the &lt;code&gt;legacy/&lt;/code&gt; folder is off-limits. Today? Blank slate. You re-explain it. Tomorrow you re-explain it again.&lt;/p&gt;

&lt;p&gt;That's not a prompt problem. It's a &lt;strong&gt;memory&lt;/strong&gt; problem.&lt;/p&gt;

&lt;h2&gt;
  
  
  Corvusmem
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://corvusmem.vercel.app/" rel="noopener noreferrer"&gt;Corvusmem&lt;/a&gt; gives your coding agent (Claude Code, Codex, or a plain CLI) a memory that actually persists — and that &lt;strong&gt;never leaves your machine&lt;/strong&gt;.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Local.&lt;/strong&gt; Your memory is a plain file in your repo. No server, no cloud, no API key, no telemetry.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Deterministic.&lt;/strong&gt; Same input, same result. No embeddings, no vector database, no "temperature".&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Zero dependencies.&lt;/strong&gt; It's small enough to read in an afternoon.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Any language.&lt;/strong&gt; Keys and values work in English, Spanish, anything — the recall is language-agnostic.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Shared, per-author.&lt;/strong&gt; On a team, everyone's facts stay separated and merge cleanly.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  30 seconds
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# in your project&lt;/span&gt;
npx corvusmem &lt;span class="nb"&gt;install&lt;/span&gt;

&lt;span class="c"&gt;# remember something&lt;/span&gt;
npx corvusmem &lt;span class="nt"&gt;--save&lt;/span&gt; &lt;span class="nt"&gt;--dimension&lt;/span&gt; project.stack &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--value&lt;/span&gt; &lt;span class="s2"&gt;"API is Fastify v5 on Node 22, PostgreSQL 16 + pgvector"&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  &lt;span class="nt"&gt;--keys&lt;/span&gt; &lt;span class="s2"&gt;"stack,backend"&lt;/span&gt;

&lt;span class="c"&gt;# recall it&lt;/span&gt;
npx corvusmem &lt;span class="nt"&gt;--recall&lt;/span&gt; &lt;span class="s2"&gt;"backend"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Once installed, the agent captures and recalls memory on its own — you don't have to babysit it. There's a hard cap on how much gets injected per message, so your context window stays small and &lt;strong&gt;doesn't grow&lt;/strong&gt; as memory does.&lt;/p&gt;

&lt;h2&gt;
  
  
  See your memory as a brain
&lt;/h2&gt;

&lt;p&gt;Run &lt;code&gt;npx corvusmem --cortex&lt;/code&gt; and it draws your whole memory as a local HTML "brain" — decisions, goals, project facts, each in its own zone. It's offline and it's oddly satisfying.&lt;/p&gt;

&lt;h2&gt;
  
  
  Honest footnote
&lt;/h2&gt;

&lt;p&gt;Corvusmem is new. I'm not going to wave fake download counts at you. It's open and free for non-commercial use (&lt;a href="https://corvusmem.vercel.app/docs" rel="noopener noreferrer"&gt;PolyForm Noncommercial&lt;/a&gt;), with a separate commercial license. If your agent forgetting everything drives you up the wall like it did me, give it a spin and tell me where it breaks.&lt;/p&gt;

&lt;p&gt;→ &lt;strong&gt;Docs &amp;amp; install:&lt;/strong&gt; &lt;a href="https://corvusmem.vercel.app/" rel="noopener noreferrer"&gt;https://corvusmem.vercel.app/&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>productivity</category>
      <category>opensource</category>
      <category>showdev</category>
    </item>
    <item>
      <title>Llegó Corvusmem: el cuervo que no olvida na'</title>
      <dc:creator>Peterson Andres Sena Arias</dc:creator>
      <pubDate>Tue, 03 Nov 2026 20:58:27 +0000</pubDate>
      <link>https://dev.to/r3v3rs3/llego-corvusmem-el-cuervo-que-no-olvida-na-15o3</link>
      <guid>https://dev.to/r3v3rs3/llego-corvusmem-el-cuervo-que-no-olvida-na-15o3</guid>
      <description>&lt;p&gt;Mira, tu agente de código es un genio… por una sesión. Después se olvida de tu stack, de lo que decidiste ayer, de quién tú ere' y de las tres cosas que le dijiste hace un rato. Cada chat nuevo arranca de cero, como si nada.&lt;/p&gt;

&lt;p&gt;A mí eso me tenía cansao de repetirle to' a Claude Code y a Codex. Así que le hice una memoria: &lt;strong&gt;&lt;a href="https://corvusmem.vercel.app" rel="noopener noreferrer"&gt;Corvusmem&lt;/a&gt;&lt;/strong&gt;. El cuervo que no olvida ni un detalle. 🐦‍⬛&lt;/p&gt;

&lt;h2&gt;
  
  
  ¿Qué es esta vaina?
&lt;/h2&gt;

&lt;p&gt;Corvusmem es una &lt;strong&gt;memoria determinista, offline y sin dependencias&lt;/strong&gt; para agentes de código. Se engancha de tu Agente' (Claude Code, Codex, o una CLI suelta), apunta los &lt;strong&gt;hechos estables&lt;/strong&gt; que tu agente aprende, y te los recuerda sesión tras sesión, proyecto tras proyecto.&lt;/p&gt;

&lt;p&gt;Y lo mejor: toda la memoria es &lt;strong&gt;un archivo que es pa' ti&lt;/strong&gt;. Sin servidor, sin nube, sin API key, sin telemetría, sin base de datos vectorial. Misma pregunta, misma respuesta — siempre.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx corvusmem &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="nt"&gt;--claude&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Ya ta'. Desde el próximo mensaje empieza a recordar.&lt;/p&gt;

&lt;h2&gt;
  
  
  Cómo funciona (sin embeddings ni brujería)
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Captura.&lt;/strong&gt; Un hook &lt;code&gt;PostToolUse&lt;/code&gt; lee cada intercambio y saca los hechos — decisiones, rutas, tu stack, tus gustos, tus objetivos.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Guarda.&lt;/strong&gt; Cada hecho cae en un &lt;strong&gt;grafo local&lt;/strong&gt; ligerito con su nombre, su fecha y sus enlaces. Nada de embeddings. Nada de base vectorial. Un archivo.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Recuerda.&lt;/strong&gt; Una consulta enciende el nodo que coincide y se extiende un salto (activación a dos saltos), y te saca lo relacionado. Al instante, y siempre igual.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Como es determinista y local, no hay nada que depurar en la nube, no hay nada que pagar, y no se va ni un byte de tu máquina.&lt;/p&gt;

&lt;h2&gt;
  
  
  Te recuerda a TI, y a tu equipo
&lt;/h2&gt;

&lt;p&gt;Hay dos tipos de memoria:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Del proyecto&lt;/strong&gt; — stack, cliente, decisiones. Vive en el repo, la comparte el equipo.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Tuya&lt;/strong&gt; — quién tú eres y cómo te gusta trabajar, en todos tus proyectos, privada en tu máquina.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Y en un proyecto compartío, Corvusmem guarda los hechos &lt;strong&gt;por autor&lt;/strong&gt;: si tú ere' de Vim y tu socio e' de Emacs, los dos sobreviven — ninguno pisa al otro. Tranquilo.&lt;/p&gt;

&lt;h2&gt;
  
  
  Y para verlo: el Cortex
&lt;/h2&gt;

&lt;p&gt;Corre &lt;code&gt;npx corvusmem --cortex&lt;/code&gt; y tu memoria se dibuja como un &lt;strong&gt;cerebro de puntos&lt;/strong&gt; — un brillo por hecho, cada tipo de memoria en su zona (objetivos, proyecto, reglas, quién ere', lo vivío). No e' adorno: e' una radiografía de lo que tu agente de verdad usa. HTML local, offline, y sigue tu tema claro/oscuro.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fkuk0q1mnea2tabzzsgjz.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fkuk0q1mnea2tabzzsgjz.png" alt="El Cortex: la memoria dibujada como un cerebro, con la línea de tiempo de lo aprendío, lo que cambió y lo que sigue abierto" width="800" height="531"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Lo honesto, sin cuento
&lt;/h2&gt;

&lt;p&gt;Te lo digo de frente: Corvusmem &lt;strong&gt;no e' el que más recuerda&lt;/strong&gt; en los benchmarks públicos. En LoCoMo anda compitiendo (~25% general) codo a codo con las memorias de embeddings — y hasta gana en saber &lt;strong&gt;decir "no sé"&lt;/strong&gt; en vez de inventar.&lt;/p&gt;

&lt;p&gt;Donde sí gana e' en lo que el benchmark no puntúa: e' &lt;strong&gt;determinista&lt;/strong&gt;, &lt;strong&gt;offline&lt;/strong&gt;, &lt;strong&gt;sin dependencias ni API keys&lt;/strong&gt;, cuesta &lt;strong&gt;cero&lt;/strong&gt;, y los datos son &lt;strong&gt;tuyos&lt;/strong&gt;. Pa' la memoria de trabajo de un agente, yo me quedo con "aburrida, privada y reproducible" antes que "unos puntitos más, pero en la nube de otro".&lt;/p&gt;

&lt;h2&gt;
  
  
  ¿Y por qué un cuervo?
&lt;/h2&gt;

&lt;p&gt;Porque el cuervo e' EL animal de la memoria — recuerda caras por año' y sabe dónde escondió cada cosa. Corvusmem = &lt;em&gt;corvus&lt;/em&gt; (cuervo) + &lt;em&gt;mem&lt;/em&gt; (memoria). El pájaro que no olvida.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpvhaw0ae69zv8ey9iqe5.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpvhaw0ae69zv8ey9iqe5.png" alt="El cuervo de Corvusmem" width="800" height="800"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Gratis, y de buena
&lt;/h2&gt;

&lt;p&gt;Corvusmem e' &lt;strong&gt;gratis pa' uso personal y no comercial&lt;/strong&gt; (PolyForm Noncommercial). Si lo va' a usar dentro de una empresa o un producto de pago, ahí sí lleva licencia comercial. Telemetría: cero, siempre.&lt;/p&gt;

&lt;h2&gt;
  
  
  Pruébalo, mano
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx corvusmem &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="nt"&gt;--claude&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;ul&gt;
&lt;li&gt;🌐 Web: &lt;a href="https://corvusmem.vercel.app" rel="noopener noreferrer"&gt;https://corvusmem.vercel.app&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;📦 npm: &lt;a href="https://www.npmjs.com/package/corvusmem" rel="noopener noreferrer"&gt;https://www.npmjs.com/package/corvusmem&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;🐙 GitHub: &lt;a href="https://github.com/petersonsenadevs/corvusmem" rel="noopener noreferrer"&gt;https://github.com/petersonsenadevs/corvusmem&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Si lo prueba', cuéntame qué recordó tu agente que te sorprendió. Y si te sirvió, déjame un ❤️ aquí abajo que eso ayuda un montón — y cualquier mejora o idea, pa'lante con ella. Comenta o abre un issue. Dale. 🐦‍⬛&lt;/p&gt;

</description>
      <category>ia</category>
      <category>opensource</category>
      <category>cli</category>
      <category>productividad</category>
    </item>
    <item>
      <title>Corvusmem: a memory for your AI coding agent that never leaves your machine</title>
      <dc:creator>Peterson Andres Sena Arias</dc:creator>
      <pubDate>Tue, 03 Nov 2026 20:57:29 +0000</pubDate>
      <link>https://dev.to/r3v3rs3/corvusmem-a-memory-for-your-ai-coding-agent-that-never-leaves-your-machine-8ke</link>
      <guid>https://dev.to/r3v3rs3/corvusmem-a-memory-for-your-ai-coding-agent-that-never-leaves-your-machine-8ke</guid>
      <description>&lt;p&gt;Your AI coding agent is brilliant for about one session. Then it forgets your stack, your decisions, who you are, and the three things you told it yesterday. Every new chat starts from zero.&lt;/p&gt;

&lt;p&gt;I got tired of re-explaining myself to Claude Code and Codex, so I built them a memory: &lt;strong&gt;&lt;a href="https://corvusmem.vercel.app" rel="noopener noreferrer"&gt;Corvusmem&lt;/a&gt;&lt;/strong&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  What it is
&lt;/h2&gt;

&lt;p&gt;Corvusmem is a &lt;strong&gt;deterministic, offline, zero-dependency memory&lt;/strong&gt; for AI coding agents. It hooks into where you already work (Claude Code, Codex, or a standalone CLI), writes down the &lt;strong&gt;stable facts&lt;/strong&gt; your agent learns, and recalls them across sessions and projects.&lt;/p&gt;

&lt;p&gt;The whole memory is a &lt;strong&gt;plain file you own&lt;/strong&gt;. No server, no cloud, no API key, no telemetry, no vector database. Same input, same result — every time.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx corvusmem &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="nt"&gt;--claude&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That's it. The hook starts capturing from the next message.&lt;/p&gt;

&lt;h2&gt;
  
  
  How it works (no embeddings, no magic)
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Capture.&lt;/strong&gt; A &lt;code&gt;PostToolUse&lt;/code&gt; hook reads each exchange and extracts facts — decisions, file paths, your stack, your preferences, goals.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Store.&lt;/strong&gt; Each fact lands in a lightweight &lt;strong&gt;local graph&lt;/strong&gt; with its name, date and links. Not embeddings. Not a vector DB. A file.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Recall.&lt;/strong&gt; A query activates the matching node and spreads one hop out (two-hop activation), surfacing the related facts. Instant, and the same answer every time.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Because it's deterministic and local, there's nothing to debug in the cloud, nothing to pay for, and nothing leaves your laptop.&lt;/p&gt;

&lt;h2&gt;
  
  
  It remembers &lt;em&gt;you&lt;/em&gt;, and your teammates
&lt;/h2&gt;

&lt;p&gt;Facts come in two kinds:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Project memory&lt;/strong&gt; — stack, client, decisions. Lives in the repo, shared by the team.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Your memory&lt;/strong&gt; — who you are and how you like to work, carried across every project, private to your machine.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;And in a shared project, Corvusmem keeps &lt;strong&gt;per-author&lt;/strong&gt; facts: if you prefer Vim and your teammate prefers Emacs, both survive — neither overwrites the other.&lt;/p&gt;

&lt;h2&gt;
  
  
  See it: the Cortex
&lt;/h2&gt;

&lt;p&gt;Run &lt;code&gt;npx corvusmem --cortex&lt;/code&gt; and your memory renders as a &lt;strong&gt;brain of points&lt;/strong&gt; — one glow per fact, each kind of memory in its own zone (goals, project, rules, who you are, what was lived). It's not a decoration; it's a readout of what your agent actually uses. Local HTML, offline, and it follows your light/dark theme.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fkuk0q1mnea2tabzzsgjz.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fkuk0q1mnea2tabzzsgjz.png" alt="The Cortex: memory rendered as an anatomical brain, with a timeline of what was learned, changed and is still open" width="800" height="531"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  The honest part
&lt;/h2&gt;

&lt;p&gt;Corvusmem is &lt;strong&gt;not&lt;/strong&gt; the highest-recall memory on the public benchmarks. On LoCoMo it's competitive (~25% overall) and trades blows with embedding-based memories — it even wins on knowing when to &lt;strong&gt;say "I don't know"&lt;/strong&gt; (abstention) instead of hallucinating.&lt;/p&gt;

&lt;p&gt;Where it actually wins is the stuff a benchmark doesn't score: it's &lt;strong&gt;deterministic&lt;/strong&gt;, &lt;strong&gt;offline&lt;/strong&gt;, &lt;strong&gt;free of dependencies and API keys&lt;/strong&gt;, costs &lt;strong&gt;nothing&lt;/strong&gt; to run, and the data is &lt;strong&gt;yours&lt;/strong&gt;. For a coding agent's working memory, I'll take "boring, private, and reproducible" over "a few points higher but in someone's cloud."&lt;/p&gt;

&lt;h2&gt;
  
  
  Why a crow?
&lt;/h2&gt;

&lt;p&gt;Crows are the memory animal — they remember faces for years and recall thousands of cached locations. Corvusmem = &lt;em&gt;corvus&lt;/em&gt; (crow) + &lt;em&gt;mem&lt;/em&gt; (memory). The bird that doesn't forget.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpvhaw0ae69zv8ey9iqe5.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpvhaw0ae69zv8ey9iqe5.png" alt="Corvusmem's crow" width="800" height="800"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Free, and fair
&lt;/h2&gt;

&lt;p&gt;Corvusmem is &lt;strong&gt;free for personal and noncommercial use&lt;/strong&gt; under PolyForm Noncommercial. Using it inside a company or a paid product needs a commercial license. Zero telemetry, always.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx corvusmem &lt;span class="nb"&gt;install&lt;/span&gt; &lt;span class="nt"&gt;--claude&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;ul&gt;
&lt;li&gt;🌐 Site: &lt;a href="https://corvusmem.vercel.app" rel="noopener noreferrer"&gt;https://corvusmem.vercel.app&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;📦 npm: &lt;a href="https://www.npmjs.com/package/corvusmem" rel="noopener noreferrer"&gt;https://www.npmjs.com/package/corvusmem&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;🐙 GitHub: &lt;a href="https://github.com/petersonsenadevs/corvusmem" rel="noopener noreferrer"&gt;https://github.com/petersonsenadevs/corvusmem&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If you try it, I'd love to hear what your agent remembered that surprised you. And if it helped you — a ❤️ on this post means a lot, and any improvement or idea is more than welcome. Drop a comment or open an issue. 🐦‍⬛&lt;/p&gt;

</description>
      <category>ai</category>
      <category>opensource</category>
      <category>cli</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Skills teach, hooks check: how my AI agent's skills and guardrails work together</title>
      <dc:creator>Peterson Andres Sena Arias</dc:creator>
      <pubDate>Mon, 12 Oct 2026 10:34:04 +0000</pubDate>
      <link>https://dev.to/r3v3rs3/skills-teach-hooks-check-how-my-ai-agents-skills-and-guardrails-work-together-nno</link>
      <guid>https://dev.to/r3v3rs3/skills-teach-hooks-check-how-my-ai-agents-skills-and-guardrails-work-together-nno</guid>
      <description>&lt;p&gt;If you've used Claude Code or Codex for a while, you've probably collected some &lt;strong&gt;skills&lt;/strong&gt;: folders of instructions that teach the agent how to do something well, like designing a UI, writing tests or auditing a backend. And maybe some &lt;strong&gt;hooks&lt;/strong&gt;: scripts that run before or after the agent acts and can stop it.&lt;/p&gt;

&lt;p&gt;Most people treat them as two separate things. A drawer of skills over here, a couple of hooks over there.&lt;/p&gt;

&lt;p&gt;That's the mistake I made for months. A skill on its own is advice: the agent reads it when it remembers to, follows it when it's convenient. A hook on its own is a wall with no map: it blocks, and the agent doesn't know what to do instead. The interesting part is what happens when &lt;strong&gt;every important skill has a hook that checks it&lt;/strong&gt;, and every hook points back to the skill that explains how to get past it.&lt;/p&gt;

&lt;p&gt;That's how &lt;strong&gt;Senzu&lt;/strong&gt; is built: 43 skills, 24 hooks and 22 commands for Claude Code and Codex. This post is a tour of the best ones and, more importantly, of how they hand work to each other.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Senzu is written in Spanish, but it works the same in English: the skills and hooks are read by the agent, and you talk to it in your language. I explained exactly how (and the three places where Spanish still leaks through) in &lt;a href="https://dev.to/r3v3rs3/my-ai-agents-guardrails-are-written-in-spanish-they-work-exactly-the-same-in-english-26p5"&gt;the previous post&lt;/a&gt;. The hook messages below are real, in Spanish, with my English translation underneath.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  The one idea: skill = how, hook = proof
&lt;/h2&gt;

&lt;p&gt;Here's the whole design in one picture:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F9z2qlwiijtiahfhg6tkk.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F9z2qlwiijtiahfhg6tkk.png" alt="Every skill that matters has a hook that checks it" width="800" height="691"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The left column teaches. The right column checks. For example:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;ui-verify&lt;/strong&gt; teaches the agent to measure a page in a real browser, phone first. &lt;strong&gt;stop-guard&lt;/strong&gt; won't let it say "done" after touching UI unless that verification happened.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;project-planner&lt;/strong&gt; teaches it to split work into small cards with a "done when" and a "how to verify". &lt;strong&gt;tarjeta-guard&lt;/strong&gt; won't let a card be marked done without the proof written in it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;depurar&lt;/strong&gt; (&lt;em&gt;debug&lt;/em&gt;) teaches a method: reproduce, write the failing test, one hypothesis, narrow it down, fix the cause. &lt;strong&gt;depurar-coach&lt;/strong&gt; fires the moment a test or build command fails and reminds the agent of that method &lt;em&gt;before&lt;/em&gt; it starts changing things at random.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If you only take one thing from this post, take that: &lt;strong&gt;for every rule you care about, write down how to do it, and then write something that checks it was done.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the hooks live: five moments of a session
&lt;/h2&gt;

&lt;p&gt;Claude Code and Codex fire hooks at specific moments. Senzu puts at least one hook in each:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fm1bjdrzuaazk77t6tc2b.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fm1bjdrzuaazk77t6tc2b.png" alt="One session, five moments, a hook in each" width="800" height="679"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;SessionStart&lt;/strong&gt;: before you've typed anything, &lt;code&gt;session-start&lt;/code&gt; tells the agent the stack, the branch, the project memory (decisions in force, what didn't work), the design system if there is one, and &lt;strong&gt;the next step of the method&lt;/strong&gt; (install, adopt, plan…).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;UserPromptSubmit&lt;/strong&gt;: when you write, &lt;code&gt;prompt-router&lt;/code&gt; matches your words against the skills and suggests the one that fits, once per skill per session. &lt;code&gt;memoria-viva&lt;/code&gt; (&lt;em&gt;living memory&lt;/em&gt;) notices when you're stating a rule ("from now on…") and makes the agent write it down.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;PreToolUse&lt;/strong&gt;: before every command or file edit. This is where the walls are: the hard no's, the method, your conventions, your architecture, your vetoes.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;PostToolUse&lt;/strong&gt;: after it acts. &lt;code&gt;edit-tracker&lt;/code&gt; records which files &lt;em&gt;this&lt;/em&gt; session touched, &lt;code&gt;format-on-save&lt;/code&gt; formats them (only if your project has a formatter configured, and never more lines than the agent changed), and &lt;code&gt;depurar-coach&lt;/code&gt; watches for failures.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Stop&lt;/strong&gt;: when the agent wants to finish. This is where "done" gets checked: verification, devlog, memory, uncommitted files.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The order matters. A hook at SessionStart can't block anything, but it can make sure the agent &lt;em&gt;starts&lt;/em&gt; in the right place. A hook at Stop can't prevent a mistake, but it can stop a mistake from being called finished.&lt;/p&gt;

&lt;p&gt;Now the best ones, in the order you'd meet them in a real feature.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. The router: read one skill, not forty
&lt;/h2&gt;

&lt;p&gt;Forty-three skills is a lot of context. If the agent loaded them all, there'd be no room left for your code. So the first pair is about &lt;strong&gt;reading less&lt;/strong&gt;:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;skill-router&lt;/strong&gt; (a skill) is a decision tree: "is this UI? backend? a bug? a plan?", and for each answer, which skill and &lt;em&gt;which section&lt;/em&gt; of it to read. Senzu's skills have a "minimum reading per task" table, so the agent opens one reference, not a manual.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;prompt-router&lt;/strong&gt; (a hook) does the first step for it. Each skill in the registry has a priority and a set of keywords. When your message matches, the hook injects one line: "this looks like a job for &lt;code&gt;gsap-scrolltrigger&lt;/code&gt;; read its SKILL.md first". The more specific skill wins: ask for a React Three Fiber scene and you get &lt;code&gt;react-three-fiber&lt;/code&gt;, not the generic &lt;code&gt;threejs-webgl&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Many of the keywords are already English (&lt;code&gt;landing&lt;/code&gt;, &lt;code&gt;hero&lt;/code&gt;, &lt;code&gt;dashboard&lt;/code&gt;, &lt;code&gt;animation&lt;/code&gt;, &lt;code&gt;debug&lt;/code&gt;, &lt;code&gt;legacy&lt;/code&gt;, &lt;code&gt;screenshot&lt;/code&gt;), which is part of why it works when you write in English.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Design before code: ui-ux-pro-max + front-skill-reminder + code-hygiene
&lt;/h2&gt;

&lt;p&gt;This trio is my favourite, because it fixed the thing agents are worst at: deciding what a page should look like &lt;em&gt;before&lt;/em&gt; writing it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;ui-ux-pro-max&lt;/strong&gt; (&lt;a href="https://github.com/nextlevelbuilder/ui-ux-pro-max-skill" rel="noopener noreferrer"&gt;by Next Level Builder&lt;/a&gt;, with a Senzu layer on top) is the entry point for anything visual. It picks style, palette, typography and layout pattern for the business and the stack, and saves them as the project's design system (&lt;code&gt;MASTER.md&lt;/code&gt;). Everything after reads from there.&lt;/p&gt;

&lt;p&gt;But a skill can be skipped. So &lt;strong&gt;front-skill-reminder&lt;/strong&gt; checks the very first UI edit of each session. If there's no design system &lt;em&gt;and&lt;/em&gt; no brief, it blocks:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fk6eit1d7896ba9yssw1h.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fk6eit1d7896ba9yssw1h.png" alt="Real output: first UI edit with no brief and no design system" width="800" height="391"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Notice the escape hatch: if it really is a one-line fix to something already built, the agent says so explicitly and retries. The wall fires once per session. It's there to force a decision, not to be annoying.&lt;/p&gt;

&lt;p&gt;Then the third piece. During the design rounds, everything you reject goes to a &lt;code&gt;gustos.md&lt;/code&gt; file (&lt;em&gt;tastes&lt;/em&gt;) under "No". &lt;strong&gt;code-hygiene&lt;/strong&gt; reads that file on every edit. Vetoed a carousel? It can't come back:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2f944dnnr05ituy5vkb8.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F2f944dnnr05ituy5vkb8.png" alt="Real output: a vetoed term is blocked" width="800" height="332"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;That's three components talking to each other through two files. The skill writes &lt;code&gt;MASTER.md&lt;/code&gt; and &lt;code&gt;gustos.md&lt;/code&gt;; one hook demands the first exists, another enforces the second. None of them would be worth much alone.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. "Done" means proven: verify-build + ui-verify + stop-guard
&lt;/h2&gt;

&lt;p&gt;This is the pair I'd install first in any project, even without the rest of Senzu.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;code-quality&lt;/strong&gt; ships a script, &lt;code&gt;verify-build.mjs&lt;/code&gt;, that figures out what kind of project it's in (Node, PHP, Python, Rust, Ruby, Java, .NET, Deno, monorepos with several of them) and runs lint, types, tests and build for what you touched. When everything passes, it leaves a small mark that says "verified after the last edit".&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;ui-verify&lt;/strong&gt; does the same for the screen: it opens the page in a real browser at 375, 768 and 1440 px, phone first, checks horizontal overflow, tap target sizes, contrast and console errors, and measures geometry in pixels. If something its container centers is 4 px off, it says so, with the probable cause. On mobile it actually opens the menu, taps it and checks it closes with Escape.&lt;/p&gt;

&lt;p&gt;And &lt;strong&gt;stop-guard&lt;/strong&gt; connects them. When the agent tries to finish, it checks: did this session edit code? Is there a verification mark &lt;em&gt;after&lt;/em&gt; the last edit? Did it touch UI? Is there a devlog entry for today? Are today's decisions in the project memory?&lt;/p&gt;

&lt;p&gt;Here's one it gave &lt;em&gt;me&lt;/em&gt;, for real, while I was writing the previous post with an agent:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fqae23ncd68q2v64710mv.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fqae23ncd68q2v64710mv.png" alt="Real output: stop-guard demanding verification, UI check and memory" width="799" height="410"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The agent had written the post, generated the images, committed, and was about to tell me "all done". It couldn't. It ran the build verification, ran the UI check at 375 and 1440 px (0 problems), added the decision to the memory, and &lt;em&gt;then&lt;/em&gt; it was done. That's the whole point: the end of a task is where agents cut corners, so that's where the strictest hook lives.&lt;/p&gt;

&lt;h2&gt;
  
  
  4. Plans that close with proof: project-planner + tarjeta-guard + arranque-guard
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;project-planner&lt;/strong&gt; turns a feature into phases of small cards. Each card says which skill to use, what "done" looks like, how to verify it, and what it's for:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight markdown"&gt;&lt;code&gt;&lt;span class="gu"&gt;### F1-T2 · Checkout page  [M] [doing]&lt;/span&gt;
&lt;span class="p"&gt;-&lt;/span&gt; What for: the customer can pay without leaving the shop
&lt;span class="p"&gt;-&lt;/span&gt; Skill: ui-ux-pro-max
&lt;span class="p"&gt;-&lt;/span&gt; Done when: a test payment ends on the thank-you page
&lt;span class="p"&gt;-&lt;/span&gt; Verify: verify-ui 375/1440 + test payment
&lt;span class="p"&gt;-&lt;/span&gt; Verified: —
&lt;span class="p"&gt;-&lt;/span&gt; Meets: —
&lt;span class="p"&gt;-&lt;/span&gt; Devlog: —
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then &lt;code&gt;/siguiente&lt;/code&gt; (&lt;em&gt;next&lt;/em&gt;) runs them one at a time. Two hooks keep that honest.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;arranque-guard&lt;/strong&gt; (&lt;em&gt;start guard&lt;/em&gt;) blocks writing code when a step of the method is missing: no Senzu installed, an existing project with no adopted conventions, or no plan and no idea what we're building, in which case the agent has to ask you first.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;tarjeta-guard&lt;/strong&gt; (&lt;em&gt;card guard&lt;/em&gt;) simulates what &lt;code&gt;PLAN.md&lt;/code&gt; would look like after the edit and, for every card that moves to done, demands the last three fields, filled with real things:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ffoldgwzje5zyxkeojal9.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ffoldgwzje5zyxkeojal9.png" alt="Real output: a card can't be marked done without proof" width="800" height="391"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;"Meets" is the subtle one. It's not enough that tests pass; the card has to say how it fulfils its &lt;em&gt;purpose&lt;/em&gt;. The tests can pass on a checkout page that doesn't let anyone pay.&lt;/p&gt;

&lt;h2&gt;
  
  
  5. Your project's rules, not the agent's: /adoptar + conventions-guard + arquitectura-guard
&lt;/h2&gt;

&lt;p&gt;On an existing codebase, the problem isn't that the agent writes bad code. It writes &lt;em&gt;different&lt;/em&gt; code: its own naming, its own validation style, its own folders.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;/adoptar&lt;/code&gt; (&lt;em&gt;adopt&lt;/em&gt;) uses the &lt;strong&gt;code-quality&lt;/strong&gt; skill to study the real project (naming, structure, validation, tests, git history, the language of identifiers and comments), proposes the conventions with a real example of each, and seals them once you agree. Then:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;conventions-guard&lt;/strong&gt; blocks code that introduces a pattern the sealed conventions forbid. Only what's &lt;em&gt;introduced&lt;/em&gt;: old code that already breaks a rule isn't your problem today.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;guard&lt;/strong&gt; makes the sealed file untouchable for the agent. It can't edit it, move it, delete it or restore it, from the editor, the terminal or a commit.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;strong&gt;ddd-hexagonal&lt;/strong&gt; is a library of 74 documents on domain-driven design, hexagonal architecture and light CQRS, including a checklist for when &lt;em&gt;not&lt;/em&gt; to use them. If you declare an architecture (MVC with services, hexagonal or DDD), &lt;strong&gt;arquitectura-guard&lt;/strong&gt; reads your real import map and blocks imports that cross a layer: the domain importing the framework, a controller using the ORM directly. I showed those messages in the previous post; the pattern is the same as everywhere else. The message says what's wrong, &lt;em&gt;where the code should go instead&lt;/em&gt;, and how you (not the agent) can approve an exception.&lt;/p&gt;

&lt;h2&gt;
  
  
  6. Memory you don't have to maintain: devlog + the memoria hooks
&lt;/h2&gt;

&lt;p&gt;The &lt;strong&gt;devlog&lt;/strong&gt; skill keeps a diary per project: one entry per relevant step, what was done, why, and how it was verified, plus a short &lt;code&gt;MEMORIA.md&lt;/code&gt; with the decisions in force (&lt;code&gt;D-012 · We do X because Y · see entry 063&lt;/code&gt;), what didn't work, and what's pending. It also has a search script that understands word stems, typos and synonyms.&lt;/p&gt;

&lt;p&gt;Writing it down is the skill. &lt;em&gt;Using&lt;/em&gt; it is the hooks':&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;session-start&lt;/code&gt; injects the memory at the start of every session. If the agent wants to contradict a decision, it has to quote it and ask.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;memoria-archivo&lt;/code&gt; (&lt;em&gt;file memory&lt;/em&gt;) fires the first time a session touches a file: "here's what the memory and the devlog say about this file". That's how the agent learns that a component broke twice for the same reason before it breaks it a third time.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;memoria-viva&lt;/code&gt; catches your rules as you say them, and &lt;code&gt;estado-sesion&lt;/code&gt; saves where the session left off, so &lt;code&gt;/retomar&lt;/code&gt; (&lt;em&gt;resume&lt;/em&gt;) can pick it up tomorrow without you explaining anything.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;stop-guard&lt;/code&gt;, again, won't let a session close with decisions in the devlog that never made it to the memory.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  7. The hard no's (that still explain themselves)
&lt;/h2&gt;

&lt;p&gt;Finally, the walls that don't pair with any skill, because there's no "right way" to do these things. You just don't:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;guard&lt;/strong&gt;: no &lt;code&gt;git push&lt;/code&gt; to main/develop/staging/production, no &lt;code&gt;rm -rf&lt;/code&gt; on dangerous paths, no destructive database commands, no &lt;code&gt;curl | bash&lt;/code&gt;, no &lt;code&gt;--force&lt;/code&gt;. The agent can't grant itself permissions or turn hooks off: only you can, from your own terminal.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;secrets-guard&lt;/strong&gt;: real-looking credentials don't go into code. AWS, GitHub, Stripe, OpenAI and Anthropic keys are recognized by their shape.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;protect-files&lt;/strong&gt;: &lt;code&gt;.env&lt;/code&gt;, already-applied migrations, lockfiles and generated files are off-limits.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;backend-guard&lt;/strong&gt;: migrations that drop or rename columns in one step, and personal data written to logs.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;cierre-limpio&lt;/strong&gt; (&lt;em&gt;clean exit&lt;/em&gt;): no ending a session with your own files uncommitted, and no touching files another agent is working on. I often run Claude Code and Codex on the same repo, and this one is what makes that possible.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fyp8zjwjkg4yndlw0464s.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fyp8zjwjkg4yndlw0464s.png" alt="Real output: secrets-guard catching a Stripe key" width="800" height="332"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What it feels like from the inside
&lt;/h2&gt;

&lt;p&gt;I'll be honest: some of these walls hit me while writing this very post.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;code-hygiene&lt;/code&gt; blocked a throwaway script I wrote to capture the hook messages, because it had a &lt;code&gt;console.log&lt;/code&gt;. Fair, and the fix was one line.&lt;/li&gt;
&lt;li&gt;The &lt;code&gt;guard&lt;/code&gt; from an older Senzu installed on my machine blocked a heredoc because the text merely &lt;em&gt;mentioned&lt;/em&gt; a protected file. That false positive is exactly what v2.16 ("quoting isn't doing") fixed: newer walls look at what actually runs, not at what a string says. I just hadn't updated that machine.&lt;/li&gt;
&lt;li&gt;And &lt;code&gt;stop-guard&lt;/code&gt;, as you saw above, didn't let the agent call the previous post finished until it was verified.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;That's the trade. A good wall costs you a few seconds now and then. A missing wall costs you the afternoon you spend finding out the agent "verified" something it never ran.&lt;/p&gt;

&lt;h2&gt;
  
  
  Build your own pairs
&lt;/h2&gt;

&lt;p&gt;You don't need Senzu to use the idea. Pick the one rule you repeat to your agent most often and write two things:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;A skill&lt;/strong&gt; that explains how to do it right: short, with a "minimum reading" section so the agent doesn't load more than it needs.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A hook&lt;/strong&gt; that checks it was done, at the right moment (PreToolUse to prevent it, Stop to refuse calling it finished), and whose message says &lt;em&gt;what to do instead&lt;/em&gt;.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;If you'd rather start from mine:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;/plugin marketplace add petersonsenadevs/senzu
/plugin install senzu-all@senzu
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then &lt;code&gt;/instalar&lt;/code&gt; in your project. In Codex, add the same marketplace from the app.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Repository (MIT): &lt;a href="https://github.com/petersonsenadevs/senzu" rel="noopener noreferrer"&gt;github.com/petersonsenadevs/senzu&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Website and docs, with a page per skill and per hook generated from the repo: &lt;a href="https://getsenzu.vercel.app" rel="noopener noreferrer"&gt;getsenzu.vercel.app&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Skills from other projects keep their author and license in &lt;code&gt;CREDITOS.md&lt;/code&gt;, and every week a job checks whether they've changed upstream and opens a PR for me to review.&lt;/p&gt;

&lt;p&gt;Which pair would you add first? I'm always looking for the wall I haven't thought of yet.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>claude</category>
      <category>opensource</category>
      <category>productivity</category>
    </item>
    <item>
      <title>My AI agent's guardrails are written in Spanish. They work exactly the same in English.</title>
      <dc:creator>Peterson Andres Sena Arias</dc:creator>
      <pubDate>Mon, 12 Oct 2026 10:13:13 +0000</pubDate>
      <link>https://dev.to/r3v3rs3/my-ai-agents-guardrails-are-written-in-spanish-they-work-exactly-the-same-in-english-26p5</link>
      <guid>https://dev.to/r3v3rs3/my-ai-agents-guardrails-are-written-in-spanish-they-work-exactly-the-same-in-english-26p5</guid>
      <description>&lt;p&gt;A couple of weeks ago I wrote (in Spanish) about &lt;strong&gt;Senzu&lt;/strong&gt;, the toolkit I built to stop repeating myself to my coding agents. Same rules in every project: don't commit to &lt;code&gt;main&lt;/code&gt;, check the page on a phone before saying it's done, write down what you did and why, pick the palette before the third component, not after.&lt;/p&gt;

&lt;p&gt;The short version of that post: rules in a &lt;code&gt;CLAUDE.md&lt;/code&gt; are suggestions. The agent reads them and &lt;em&gt;almost always&lt;/em&gt; follows them. "Almost always" isn't good enough for a real project, so I turned every rule that could be executed into something that executes: hooks that block, not paragraphs that advise.&lt;/p&gt;

&lt;p&gt;Since then Senzu has gone from v2.7 to v2.19. That's a lot of new walls, a guided start for new and existing projects, a memory that keeps itself up to date, and a website. And the question I got the most after that first post was not about any of that. It was this:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;"It's all in Spanish. Does it work if I don't speak Spanish?"&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Yes. This post explains why, and then shows you everything that's new.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Spanish, and why it doesn't matter
&lt;/h2&gt;

&lt;p&gt;Senzu is written in Spanish on purpose. There's a mountain of tooling for coding agents in English and almost nothing designed for people who work in Spanish, so I built the thing I wanted to use.&lt;/p&gt;

&lt;p&gt;But here's what took me a while to see clearly: &lt;strong&gt;almost nothing in Senzu is written for you. It's written for the agent.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F71qcfj7g5ld7f2d5eorp.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F71qcfj7g5ld7f2d5eorp.png" alt="Who reads what: Spanish for the agent, your language for you"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Think about what Senzu actually is:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Skills&lt;/strong&gt; are instructions the agent loads when a task needs them: how to design a UI, how to audit a backend, how to plan a feature. The agent reads them. You never do, unless you're curious.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Hooks&lt;/strong&gt; ("walls", &lt;em&gt;muros&lt;/em&gt; in Spanish) are small Node scripts that run before and after the agent acts. They inspect the command or the file and either let it through or block it with a reason. The reason goes to the agent.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Commands&lt;/strong&gt; like &lt;code&gt;/plan&lt;/code&gt; or &lt;code&gt;/adoptar&lt;/code&gt; are prompts with a name. You type the name; the agent reads the prompt.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Claude and GPT read Spanish as well as they read English. An instruction like "el dominio no conoce el framework" lands exactly the same as "the domain doesn't know about the framework". The model doesn't translate it into a weaker rule, it just understands it.&lt;/p&gt;

&lt;p&gt;What &lt;em&gt;you&lt;/em&gt; see is the conversation, and the conversation happens in your language. The hook that runs at the start of every session tells the agent, literally: &lt;em&gt;"respond in Spanish, or in the language the user writes to you in"&lt;/em&gt;. Write in English, get English back.&lt;/p&gt;

&lt;p&gt;Your code doesn't become Spanish either. When you adopt an existing project, Senzu learns its conventions, including the language of its identifiers and comments, and seals them (more on that below). An English codebase stays English.&lt;/p&gt;

&lt;p&gt;There are three places where Spanish leaks through, and I'd rather tell you than have you find out:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;The rules file installed in your project&lt;/strong&gt; (&lt;code&gt;CLAUDE.md&lt;/code&gt; / &lt;code&gt;AGENTS.md&lt;/code&gt;) opens with &lt;em&gt;"Comunícate en español"&lt;/em&gt;. The session hook overrides it with "or the user's language", but if you want to remove any doubt, tell Senzu once and it will remember for every project:
&lt;/li&gt;
&lt;/ol&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;   /recordar always answer me in English, and write the devlog and plan in English
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That line goes to your &lt;em&gt;user memory&lt;/em&gt; (&lt;code&gt;~/.config/senzu/memoria.md&lt;/code&gt;), which is injected at the start of every session in every project, with the instruction that it beats the agent's own habits.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;When a wall blocks something, the raw message is in Spanish.&lt;/strong&gt; You'll see it scroll by in the tool output. The agent reads it, understands it, and explains it to you in English. Here's a real one, so you know what to expect:&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F1v61lerweara2q3gbwc6.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F1v61lerweara2q3gbwc6.png" alt="A real wall: no commits on main, with the English translation"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Real output of the &lt;code&gt;guard&lt;/code&gt; hook. The "In English" box is my translation for this post; in a session, the agent explains it to you.&lt;/em&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Automatic capture of your corrections listens for Spanish.&lt;/strong&gt; One of the new features (below) notices when you say things like &lt;em&gt;"no vuelvas a…"&lt;/em&gt; ("don't do that again") and makes the agent write it down. The phrases it listens for are Spanish, so "from now on, never…" in English slips past it. The workaround is the same command: when you want something remembered, say &lt;code&gt;/recordar …&lt;/code&gt; and it's stored, in any language. English phrases for the detector are on my list.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;That's it. Everything else is the same as if Senzu were written in English: the same walls fire, the same commands run, the same skills load.&lt;/p&gt;
&lt;h2&gt;
  
  
  Quick recap: what Senzu is
&lt;/h2&gt;

&lt;p&gt;If you didn't read the first post:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;A package for &lt;strong&gt;Claude Code and Codex&lt;/strong&gt; (Codex runs the same hooks since v1.5).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;43 skills&lt;/strong&gt;: mine, plus the best ones I found from other people (with their author and license credited), with a layer on top that makes them work together. A router decides which one to load, so the context isn't flooded with things you don't need.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;24 hooks&lt;/strong&gt; that block or remind.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;22 commands&lt;/strong&gt; for the workflows I repeat in every project.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;9 stacks&lt;/strong&gt; with their own rules: Laravel, Next, Nuxt, Astro, SvelteKit, Vue + TypeScript, Node APIs, WordPress and Python with LangGraph.&lt;/li&gt;
&lt;li&gt;MIT licensed.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Now, what's new.&lt;/p&gt;
&lt;h2&gt;
  
  
  New: a guided start, for new projects &lt;em&gt;and&lt;/em&gt; inherited ones
&lt;/h2&gt;

&lt;p&gt;The biggest change since the first post is that Senzu now has an opinion about &lt;strong&gt;the order of things&lt;/strong&gt;, and enforces it.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fh03vmpisthidgv84zu45.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fh03vmpisthidgv84zu45.png" alt="The method: one command per step, and a wall if you skip one"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;When you open a session, the start hook figures out where you are and tells the agent the next step of the method:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Empty or new project:&lt;/strong&gt; &lt;code&gt;/brief&lt;/code&gt; (a plain-language interview: what the business does, what visitors should do, what sites you like), then &lt;code&gt;/propuestas&lt;/code&gt; (two genuinely different mockups you can open in a browser and vote on piece by piece), &lt;code&gt;/design-system&lt;/code&gt;, &lt;code&gt;/plan&lt;/code&gt;, and &lt;code&gt;/siguiente&lt;/code&gt; to execute one task at a time.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Existing project with code but no sealed conventions:&lt;/strong&gt; &lt;code&gt;/adoptar&lt;/code&gt; first. Before the agent writes a single line, it studies how the project is &lt;em&gt;actually&lt;/em&gt; written and agrees the rules with you.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;And it's not a suggestion. The &lt;code&gt;arranque-guard&lt;/code&gt; hook blocks the first attempt to write code if the missing step hasn't happened. No plan? The agent has to ask you what you're building before it starts.&lt;/p&gt;

&lt;p&gt;The new website has a section for exactly this:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ff3ycrlelh09ag2m4l3du.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Ff3ycrlelh09ag2m4l3du.png" alt="The two paths on the Senzu website"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Each step links to its command page. The descriptions come straight from the Senzu repo at build time, so they never drift.&lt;/em&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  &lt;code&gt;/adoptar&lt;/code&gt;: the agent writes like your project, not like itself
&lt;/h2&gt;

&lt;p&gt;This is the one I'm proudest of, because it fixes the thing that annoyed me most.&lt;/p&gt;

&lt;p&gt;Drop an agent into a three-year-old Laravel project and it will happily write code in &lt;em&gt;its&lt;/em&gt; style: different naming, a different validation approach, a different folder layout. Every PR becomes a style argument.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;/adoptar&lt;/code&gt; (&lt;em&gt;adopt&lt;/em&gt;) reads the real code: naming, structure, how validation is done, how tests are named, what the git history looks like, which language identifiers and comments are in. It proposes the conventions, with a real example from your code for each one. You agree or correct. Then it &lt;strong&gt;seals&lt;/strong&gt; them in &lt;code&gt;senzu/conventions.md&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;From then on:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;The &lt;code&gt;conventions-guard&lt;/code&gt; hook blocks code that breaks the sealed conventions.&lt;/li&gt;
&lt;li&gt;The agent &lt;strong&gt;can't edit, move or delete&lt;/strong&gt; the sealed file, not with an edit, not from the terminal, not by sneaking it into a commit. In v2.14 I found it could be bypassed if the path came in a different format (&lt;code&gt;C:/&lt;/code&gt; vs &lt;code&gt;/c/&lt;/code&gt;) and fixed it as a security issue.&lt;/li&gt;
&lt;li&gt;The project's convention beats Senzu's own preferences. Always.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2&gt;
  
  
  New: declared architecture, enforced by a wall
&lt;/h2&gt;

&lt;p&gt;v2.18 adds something I wanted for a long time. You declare your architecture once: &lt;strong&gt;MVC with services, hexagonal, or DDD&lt;/strong&gt;. There are templates for Laravel, Node and Python. A wall makes the code respect it.&lt;/p&gt;

&lt;p&gt;It checks layers, bounded contexts, controllers that talk to the ORM directly, DTOs and folders. It reads the real import map of your project (PSR-4, &lt;code&gt;tsconfig&lt;/code&gt; aliases, Python modules), so it knows that &lt;code&gt;App\Models\Order&lt;/code&gt; lives in infrastructure.&lt;/p&gt;

&lt;p&gt;Here's what happens when the agent tries to put Eloquent in a controller of a hexagonal project:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F9q31g661xcyk3qjh4zz3.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F9q31g661xcyk3qjh4zz3.png" alt="A real wall: the controller can't use infrastructure"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;And when it tries to import the framework into the domain:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fglajkd2r7r7282j0mmmm.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fglajkd2r7r7282j0mmmm.png" alt="A real wall: the domain can't import the framework"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Both are real outputs, produced by running the hook against a throwaway Laravel project for this post.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Notice the last line of each message: it doesn't just say no, it says &lt;em&gt;where the code should go instead&lt;/em&gt;. That's what turns a wall into something the agent can act on in its next attempt. If you really need an exception, you approve it and it's written down with a reason, right on that line.&lt;/p&gt;
&lt;h2&gt;
  
  
  New: a feature isn't done until the backend says so
&lt;/h2&gt;

&lt;p&gt;v2.17 extended the "no 'done' without proof" rule from the frontend to the backend. The &lt;code&gt;feature-guard&lt;/code&gt; hook won't let the agent close a backend feature if:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;there's no test for it,&lt;/li&gt;
&lt;li&gt;there's a migration that was never run, or one without a working &lt;code&gt;down&lt;/code&gt;,&lt;/li&gt;
&lt;li&gt;a new config key was added but not to &lt;code&gt;.env.example&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;And every command now ends by saying what the next step is, so the chain from &lt;code&gt;/brief&lt;/code&gt; to a deployed feature doesn't break in the middle.&lt;/p&gt;

&lt;p&gt;Related: plan cards can't be marked as done without evidence (&lt;code&gt;tarjeta-guard&lt;/code&gt;), and when the agent says it verified something, verifying means real output: lint, types, tests, build, and the mobile check for UI work. v2.19 taught the build verifier more languages (Rust, Ruby, Java, .NET, Deno, Makefiles) and what to do in projects with no manifest at all, instead of asking for a fake verification.&lt;/p&gt;
&lt;h2&gt;
  
  
  New: memory that maintains itself
&lt;/h2&gt;

&lt;p&gt;The first post described the project memory: a short file with the decisions in force (&lt;code&gt;D-012 · We do X because Y · see entry 063&lt;/code&gt;), injected at the start of every session. If the agent wants to contradict a decision, it has to quote it and ask.&lt;/p&gt;

&lt;p&gt;v2.11 made it alive:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Your corrections are written down automatically.&lt;/strong&gt; When you correct the agent with a rule ("don't do that again", "from now on…"), the &lt;code&gt;memoria-viva&lt;/code&gt; hook notices and makes the agent write it down before going on; if it forgets, it can't close the session. (As mentioned above, it listens for the Spanish phrasing for now; in English, use &lt;code&gt;/recordar&lt;/code&gt;.)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;User memory&lt;/strong&gt;: rules that apply to all your projects (language, git habits, how you like to be spoken to) live in one file outside any project. That's the &lt;code&gt;/recordar&lt;/code&gt; trick from the language section.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Per-file memory&lt;/strong&gt;: when the agent opens a file that has history ("this file broke twice because of X"), it gets reminded.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;&lt;code&gt;/retomar&lt;/code&gt;&lt;/strong&gt; (&lt;em&gt;resume&lt;/em&gt;): picks up where the last session left off, with what was done, what's pending and what was decided.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2&gt;
  
  
  New: two agents, one repo, no mess
&lt;/h2&gt;

&lt;p&gt;I often have Claude Code and Codex working on the same repository. That used to end badly: one agent committing the other's half-finished files, or finishing its session with a pile of uncommitted changes.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Clean exit&lt;/strong&gt; (&lt;code&gt;cierre-limpio&lt;/code&gt;, v2.10): the agent can't end a session leaving its own files uncommitted, and it can't touch files another agent is working on.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Protected branches&lt;/strong&gt; (v2.12, refined in v2.19.2): no commits or pushes to &lt;code&gt;main&lt;/code&gt;, &lt;code&gt;develop&lt;/code&gt;, &lt;code&gt;staging&lt;/code&gt; or &lt;code&gt;production&lt;/code&gt;. Pushing to your own work branch (&lt;code&gt;feat/…&lt;/code&gt;, &lt;code&gt;fix/…&lt;/code&gt;) is allowed out of the box.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;"Quoting isn't doing"&lt;/strong&gt; (v2.16): walls used to fire when a commit message, a heredoc, an &lt;code&gt;echo&lt;/code&gt;, a &lt;code&gt;grep&lt;/code&gt; or a comment merely &lt;em&gt;mentioned&lt;/em&gt; something forbidden. Now they look at what actually runs. Fewer false positives means fewer reasons to turn a wall off.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2&gt;
  
  
  New: install profiles, and it tells you when it's out of date
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Profiles&lt;/strong&gt; (v2.9): instead of "everything or pick one by one", &lt;code&gt;/instalar&lt;/code&gt; offers profiles: web, backend, frontend, AI agent, library. Each one brings the skills, walls and commands that make sense for that kind of project.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Out-of-date warning&lt;/strong&gt; (v2.15): at the start of a session, if your installed Senzu is older than the latest release, the agent tells you, along with the two or three changes that matter most to you. No changelog reading required.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2&gt;
  
  
  New: a website, with docs generated from the repo
&lt;/h2&gt;

&lt;p&gt;Senzu has a home now: &lt;a href="https://getsenzu.vercel.app" rel="noopener noreferrer"&gt;getsenzu.vercel.app&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fqrftnpbgcxqxnzonaytx.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fqrftnpbgcxqxnzonaytx.png" alt="The Senzu website"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The part I like most is invisible. &lt;strong&gt;The documentation is never written by hand.&lt;/strong&gt; On every build, the site fetches the latest tagged release of the Senzu repo and generates a page per skill, per command and per hook from the real files. The numbers on the home page (skills, hooks, commands, stacks) come from the same place. If a hook is added upstream, it shows up on the next deploy; if a command is removed, its page disappears.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F7nmh1l4nj7rnlnpip6af.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2F7nmh1l4nj7rnlnpip6af.png" alt="Every hook has its own page, generated from the repo"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The site and its docs are in Spanish too. If you're reading in English, your browser's translate button does a good job, and so does the agent itself: ask it "what does the &lt;code&gt;feature-guard&lt;/code&gt; hook do?" and it'll read the docs and answer in English.&lt;/p&gt;
&lt;h2&gt;
  
  
  How to try it
&lt;/h2&gt;

&lt;p&gt;In Claude Code:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;/plugin marketplace add petersonsenadevs/senzu
/plugin install senzu-all@senzu
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;In Codex, add the same marketplace from the app. Then, inside your project, open a session and type:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;/instalar
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;It detects the stack, asks what you want (a profile, everything, or à la carte) and sets it up. Then, if you want everything in English:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;/recordar always answer me in English, and write the devlog and plan in English
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;From there, if the project already has code, start with &lt;code&gt;/adoptar&lt;/code&gt;. If it's a new one, &lt;code&gt;/brief&lt;/code&gt;. The agent will tell you anyway.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Repository: &lt;a href="https://github.com/petersonsenadevs/senzu" rel="noopener noreferrer"&gt;github.com/petersonsenadevs/senzu&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Website and docs: &lt;a href="https://getsenzu.vercel.app" rel="noopener noreferrer"&gt;getsenzu.vercel.app&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;The first post, in Spanish: &lt;a href="https://dev.to/r3v3rs3/le-puse-muros-a-mi-agente-de-ia-y-ahora-rompe-la-web-pero-porque-se-lo-pido-3i8b"&gt;Le puse muros a mi agente de IA&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What it isn't
&lt;/h2&gt;

&lt;p&gt;It's &lt;strong&gt;opinionated&lt;/strong&gt;. These are my working rules, not everyone's. You can turn walls off per project and change permissions, but only you can: the agent can't switch off its own walls, and that's by design.&lt;/p&gt;

&lt;p&gt;It &lt;strong&gt;doesn't make the agent stop making mistakes&lt;/strong&gt;. It still makes them. The difference is that now it finds out before I do, and the explanation of what to do instead arrives in the same message.&lt;/p&gt;

&lt;p&gt;And it's &lt;strong&gt;in Spanish&lt;/strong&gt;, which, as it turns out, matters a lot less than I thought. The agent doesn't care. Hopefully you won't either.&lt;/p&gt;

&lt;p&gt;If you try it, tell me which wall you'd add. There's always one I missed.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>claude</category>
      <category>opensource</category>
      <category>webdev</category>
    </item>
    <item>
      <title>[Boost]</title>
      <dc:creator>Peterson Andres Sena Arias</dc:creator>
      <pubDate>Sat, 03 Oct 2026 15:19:31 +0000</pubDate>
      <link>https://dev.to/r3v3rs3/-4458</link>
      <guid>https://dev.to/r3v3rs3/-4458</guid>
      <description>&lt;div class="ltag__link--embedded"&gt;
  &lt;div class="crayons-story "&gt;
  &lt;a href="https://dev.to/r3v3rs3/le-puse-muros-a-mi-agente-de-ia-y-ahora-rompe-la-web-pero-porque-se-lo-pido-3i8b" class="crayons-story__hidden-navigation-link"&gt;Le puse muros a mi agente de IA (y ahora rompe la web, pero porque se lo pido)&lt;/a&gt;


  &lt;div class="crayons-story__body crayons-story__body-full_post"&gt;
    &lt;div class="crayons-story__top"&gt;
      &lt;div class="crayons-story__meta"&gt;
        &lt;div class="crayons-story__author-pic"&gt;

          &lt;a href="/r3v3rs3" class="crayons-avatar  crayons-avatar--l  "&gt;
            &lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F1249042%2Fe1afca32-f78d-47dd-8926-3c2c75df7c61.jpeg" alt="r3v3rs3 profile" class="crayons-avatar__image" width="460" height="460"&gt;
          &lt;/a&gt;
        &lt;/div&gt;
        &lt;div&gt;
          &lt;div&gt;
            &lt;a href="/r3v3rs3" class="crayons-story__secondary fw-medium m:hidden"&gt;
              Peterson Andres Sena Arias
            &lt;/a&gt;
            &lt;div class="profile-preview-card relative mb-4 s:mb-0 fw-medium hidden m:inline-block"&gt;
              
                Peterson Andres Sena Arias
                &lt;a href="/++"&gt;&lt;img alt="Subscriber" class="subscription-icon" src="https://assets.dev.to/assets/subscription-icon-805dfa7ac7dd660f07ed8d654877270825b07a92a03841aa99a1093bd00431b2.png" width="166" height="102"&gt;&lt;/a&gt;
                
              
              &lt;div id="story-author-preview-content-4791692" class="profile-preview-card__content crayons-dropdown branded-7 p-4 pt-0"&gt;
                &lt;div class="gap-4 grid"&gt;
                  &lt;div class="-mt-4"&gt;
                    &lt;a href="/r3v3rs3" class="flex"&gt;
                      &lt;span class="crayons-avatar crayons-avatar--xl mr-2 shrink-0"&gt;
                        &lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F1249042%2Fe1afca32-f78d-47dd-8926-3c2c75df7c61.jpeg" class="crayons-avatar__image" alt="" width="460" height="460"&gt;
                      &lt;/span&gt;
                      &lt;span class="crayons-link crayons-subtitle-2 mt-5"&gt;Peterson Andres Sena Arias&lt;/span&gt;
                    &lt;/a&gt;
                  &lt;/div&gt;
                  &lt;div class="print-hidden"&gt;
                    
                      Follow
                    
                  &lt;/div&gt;
                  &lt;div class="author-preview-metadata-container"&gt;&lt;/div&gt;
                &lt;/div&gt;
              &lt;/div&gt;
            &lt;/div&gt;

          &lt;/div&gt;
          &lt;a href="https://dev.to/r3v3rs3/le-puse-muros-a-mi-agente-de-ia-y-ahora-rompe-la-web-pero-porque-se-lo-pido-3i8b" class="crayons-story__tertiary fs-xs"&gt;&lt;time&gt;Oct 3&lt;/time&gt;&lt;span class="time-ago-indicator-initial-placeholder"&gt;&lt;/span&gt;&lt;/a&gt;
        &lt;/div&gt;
      &lt;/div&gt;

    &lt;/div&gt;

    &lt;div class="crayons-story__indention"&gt;
      &lt;h2 class="crayons-story__title crayons-story__title-full_post"&gt;
        &lt;a href="https://dev.to/r3v3rs3/le-puse-muros-a-mi-agente-de-ia-y-ahora-rompe-la-web-pero-porque-se-lo-pido-3i8b" id="article-link-4791692"&gt;
          Le puse muros a mi agente de IA (y ahora rompe la web, pero porque se lo pido)
        &lt;/a&gt;
      &lt;/h2&gt;
        &lt;div class="crayons-story__tags"&gt;
            &lt;a class="crayons-tag  crayons-tag--monochrome " href="/t/spanish"&gt;&lt;span class="crayons-tag__prefix"&gt;#&lt;/span&gt;spanish&lt;/a&gt;
            &lt;a class="crayons-tag  crayons-tag--monochrome " href="/t/ai"&gt;&lt;span class="crayons-tag__prefix"&gt;#&lt;/span&gt;ai&lt;/a&gt;
            &lt;a class="crayons-tag  crayons-tag--monochrome " href="/t/claude"&gt;&lt;span class="crayons-tag__prefix"&gt;#&lt;/span&gt;claude&lt;/a&gt;
            &lt;a class="crayons-tag  crayons-tag--monochrome " href="/t/codex"&gt;&lt;span class="crayons-tag__prefix"&gt;#&lt;/span&gt;codex&lt;/a&gt;
        &lt;/div&gt;
      &lt;div class="crayons-story__bottom"&gt;
        &lt;div class="crayons-story__details"&gt;
          &lt;a href="https://dev.to/r3v3rs3/le-puse-muros-a-mi-agente-de-ia-y-ahora-rompe-la-web-pero-porque-se-lo-pido-3i8b" class="crayons-btn crayons-btn--s crayons-btn--ghost crayons-btn--icon-left"&gt;
            &lt;div class="multiple_reactions_aggregate"&gt;
              &lt;span class="multiple_reactions_icons_container"&gt;
                  &lt;span class="crayons_icon_container"&gt;
                    &lt;img src="https://assets.dev.to/assets/sparkle-heart-5f9bee3767e18deb1bb725290cb151c25234768a0e9a2bd39370c382d02920cf.svg" width="24" height="24"&gt;
                  &lt;/span&gt;
              &lt;/span&gt;
              &lt;span class="aggregate_reactions_counter"&gt;1&lt;span class="hidden s:inline"&gt;&amp;nbsp;reaction&lt;/span&gt;&lt;/span&gt;
            &lt;/div&gt;
          &lt;/a&gt;
            &lt;a href="https://dev.to/r3v3rs3/le-puse-muros-a-mi-agente-de-ia-y-ahora-rompe-la-web-pero-porque-se-lo-pido-3i8b#comments" class="crayons-btn crayons-btn--s crayons-btn--ghost crayons-btn--icon-left flex items-center"&gt;
              

              &lt;span class="hidden s:inline"&gt;Add&amp;nbsp;Comment&lt;/span&gt;
            &lt;/a&gt;
        &lt;/div&gt;
        &lt;div class="crayons-story__save"&gt;
          &lt;small class="crayons-story__tertiary fs-xs mr-2"&gt;
            9 min read
          &lt;/small&gt;
        &lt;/div&gt;
      &lt;/div&gt;
    &lt;/div&gt;
  &lt;/div&gt;
&lt;/div&gt;

&lt;/div&gt;


</description>
      <category>agents</category>
      <category>ai</category>
      <category>llm</category>
    </item>
  </channel>
</rss>
