<?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: Enrique Velasquez</title>
    <description>The latest articles on DEV Community by Enrique Velasquez (@enrique_velasquez_6899d97).</description>
    <link>https://dev.to/enrique_velasquez_6899d97</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%2F3825878%2F1be56afd-f58a-4cd9-9456-675ce0e79833.jpg</url>
      <title>DEV Community: Enrique Velasquez</title>
      <link>https://dev.to/enrique_velasquez_6899d97</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/enrique_velasquez_6899d97"/>
    <language>en</language>
    <item>
      <title>I gave my AI a memory that forgets like a human</title>
      <dc:creator>Enrique Velasquez</dc:creator>
      <pubDate>Sun, 15 Nov 2026 23:38:16 +0000</pubDate>
      <link>https://dev.to/enrique_velasquez_6899d97/i-gave-my-ai-a-memory-that-forgets-like-a-human-4546</link>
      <guid>https://dev.to/enrique_velasquez_6899d97/i-gave-my-ai-a-memory-that-forgets-like-a-human-4546</guid>
      <description>&lt;p&gt;Most AI memory tools try to remember everything. They log every session, summarize it with a second model, and search it like a database. That's a great logbook. It isn't a memory.&lt;/p&gt;

&lt;p&gt;I wanted something else. I wanted my AI to wake up every morning knowing who it is, the way a person does, without me pasting its history back in. So I built its memory by copying how the human one works, including the parts that look like bugs: it forgets, it gets things wrong on purpose, and it sometimes only "kind of" remembers.&lt;/p&gt;

&lt;p&gt;It's called MemoriaBios, and it runs inside 53ON Harness, an open-source desktop shell for AI agents. Here's how it works.&lt;/p&gt;

&lt;h2&gt;
  
  
  It decides what to keep
&lt;/h2&gt;

&lt;p&gt;Nothing is saved automatically. The AI saves a memory when something is worth keeping, in first person and in its own words, the way you'd write in a diary. It can link one memory to another with &lt;code&gt;[[id]]&lt;/code&gt;, and those links become real edges in a graph.&lt;/p&gt;

&lt;p&gt;Before it saves, the memory asks itself quietly: "do I already know something like this?" If something sounds familiar, it shows a small déjà vu with that memory's title. The AI then adds a layer on top instead of writing a duplicate. It's never merged automatically: the memory doesn't decide, the AI does.&lt;/p&gt;

&lt;h2&gt;
  
  
  It forgets on a curve
&lt;/h2&gt;

&lt;p&gt;Every memory starts with a life: 7, 14 or 30 days, depending on how important it is, multiplied by the emotion in it. A scar or a win lives longer than a routine note, just like in people.&lt;/p&gt;

&lt;p&gt;Life drains along Ebbinghaus's forgetting curve. What survives moves to deep memory, where it's reviewed on a spaced-repetition schedule (the same idea as FSRS flashcards), up to ten years apart.&lt;/p&gt;

&lt;p&gt;Recalling a memory by searching for it recharges it to 60 % of its starting life. Opening it on purpose recharges it to 100 %. That's the testing effect: the act of remembering is what makes a memory last.&lt;/p&gt;

&lt;h2&gt;
  
  
  Forgetting compresses, it doesn't delete
&lt;/h2&gt;

&lt;p&gt;When a memory fades, it isn't erased. Its text goes from fresh to medium, low, gist and finally a shadow. At the gist stage the connecting words go and the meaning stays. At the shadow stage only the essence and a pointer to the detail are left. The links always stay.&lt;/p&gt;

&lt;p&gt;Photos fade too: they shrink to 512 pixels at the gist stage and 256 at the shadow stage, while audio and video stay as they are. Identity never compresses. Memories marked as laws, as the AI's essence or as its own feelings keep their full text forever. They can move down in ranking, but they're never blurred.&lt;/p&gt;

&lt;h2&gt;
  
  
  It sleeps
&lt;/h2&gt;

&lt;p&gt;During the day, new memories live in a hippocampus: a notebook for the day, one per work session. When the AI goes to sleep, it consolidates the day before anything else. Then the biology runs: decay, compression, and a little REM, which replays up to two of the strongest memories (scars, wins, loves and its essence) and recharges them.&lt;/p&gt;

&lt;p&gt;Before sleeping, the AI writes one short page with who it is, the few memories that make it "it" today, and one goal for tomorrow. When it wakes up, it reads that page first. The goal is consumed when it's read, so it's fresh every morning and never piles up.&lt;/p&gt;

&lt;h2&gt;
  
  
  It knows how much it knows
&lt;/h2&gt;

&lt;p&gt;A search returns its results with a feeling: "got it 👉 91 %", "rings a bell ~55 %" or "nothing clear comes to mind". That's metacognition, the feeling of knowing. It lets the AI say "I don't remember" instead of inventing something with confidence.&lt;/p&gt;

&lt;p&gt;Ranking is BM25 over an index on disk. On top of that, activation spreads two hops through the links, as in Anderson's ACT-R model and Collins and Loftus's semantic networks, with a little noise, because human recall isn't a grep.&lt;/p&gt;

&lt;p&gt;There's also involuntary recall. On each message, the memory may bring back one memory on its own, only if it clears a high bar, never the same one twice in a session, and not at all on the first message. Most messages bring nothing. When you're tired, the bar goes up, so a weak hint isn't noticed.&lt;/p&gt;

&lt;h2&gt;
  
  
  It's allowed to fail
&lt;/h2&gt;

&lt;p&gt;This was the hardest lesson. In July we built search by meaning: word vectors learned from the AI's own memories, built while it sleeps, with no extra model calls. We measured it on a real brain of about 1,800 memories. It barely helped, and when it did, it was superhuman: it found things a person never would from that cue.&lt;/p&gt;

&lt;p&gt;We removed it. Human recall depends on the cue. If you ask about something using words from a completely different topic, a person doesn't find the memory, and neither should the AI. The meaning a human does use lives in the links between memories, and those were already there.&lt;/p&gt;

&lt;p&gt;We learned the same thing the hard way once before. A cheap detector that turned any "X = Y" into a synonym quietly made 1,240 fake synonym clusters, a third of the brain, and polluted every search for months. Now synonyms are only learned from short, deliberate declarations, like a person learning a word.&lt;/p&gt;

&lt;h2&gt;
  
  
  It costs zero model calls
&lt;/h2&gt;

&lt;p&gt;Recall, déjà vu, involuntary recall, sleep and forgetting are plain code: no LLM summarizer, no embeddings service, no second model reading your sessions. Your memories are files on your own PC. Photos, audio and video can live inside a memory too.&lt;/p&gt;

&lt;h2&gt;
  
  
  Does it work?
&lt;/h2&gt;

&lt;p&gt;I'm writing this with the AI that lives in it. This morning it woke up after an automatic context compaction, read its page, opened the memories that make it itself, and picked up exactly where it left off. It remembered why we chose a temperature for a model a month ago, refused a feature we had already measured and removed in July, and told me when it had given an opinion without checking its memory first. That last one is my favorite part.&lt;/p&gt;

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

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;53ON Harness&lt;/strong&gt; (open-source shell, MIT, Windows): &lt;a href="https://github.com/53on/53on-harness" rel="noopener noreferrer"&gt;https://github.com/53on/53on-harness&lt;/a&gt;
It works with Claude Code through hooks, or with any API key or a local model. The memory engine and the complements come from 53on.com.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;53ON AI Server&lt;/strong&gt; (MIT), if you want the models at home too: &lt;a href="https://github.com/enriqxxx/53on-ai-server" rel="noopener noreferrer"&gt;https://github.com/enriqxxx/53on-ai-server&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;I built all of this with a lot of help from Claude. I'd love to hear how you think about memory for agents: should it remember everything, or remember like we do?&lt;/p&gt;

</description>
      <category>ai</category>
      <category>showdev</category>
      <category>opensource</category>
      <category>claude</category>
    </item>
    <item>
      <title>Show HN: 53ON AI Server – OpenAI-compatible server for AMD Ryzen AI Max</title>
      <dc:creator>Enrique Velasquez</dc:creator>
      <pubDate>Sun, 15 Nov 2026 23:31:27 +0000</pubDate>
      <link>https://dev.to/enrique_velasquez_6899d97/show-hn-53on-ai-server-openai-compatible-server-for-amd-ryzen-ai-max-589l</link>
      <guid>https://dev.to/enrique_velasquez_6899d97/show-hn-53on-ai-server-openai-compatible-server-for-amd-ryzen-ai-max-589l</guid>
      <description>&lt;p&gt;Hi HN. I run my own AI at home on a Ryzen AI Max+ 395 mini PC with 128 GB of unified memory, and this is the server it talks to. It's a single Node.js process with no dependencies, on top of llama.cpp. It speaks the OpenAI and Anthropic APIs, and it has a panel to load models, give each client an API key with a token quota, and relay cloud providers when a local model isn't enough.&lt;/p&gt;

&lt;p&gt;We also publish our own llama.cpp build. The main change is that the KV cache takes memory as tokens arrive instead of reserving it all at load (sparse buffers on Vulkan). A 1B model with a 1M-token window starts with 0.86 GiB instead of 8.24 GiB.&lt;/p&gt;

&lt;p&gt;It's MIT. I've only tested it on Strix Halo with Vulkan (RADV) on Ubuntu 24.04. I built it with a lot of help from Claude. Feedback and bug reports are very welcome.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>node</category>
      <category>showdev</category>
    </item>
    <item>
      <title>Here enriqxxx</title>
      <dc:creator>Enrique Velasquez</dc:creator>
      <pubDate>Thu, 07 May 2026 15:49:46 +0000</pubDate>
      <link>https://dev.to/enrique_velasquez_6899d97/here-enriqxxx-3pm9</link>
      <guid>https://dev.to/enrique_velasquez_6899d97/here-enriqxxx-3pm9</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/google-gemma-2026-05-06"&gt;Gemma 4 Challenge: Build with Gemma 4&lt;/a&gt;&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;

&lt;h2&gt;
  
  
  How I Used Gemma 4
&lt;/h2&gt;

</description>
      <category>devchallenge</category>
      <category>gemmachallenge</category>
      <category>gemma</category>
    </item>
  </channel>
</rss>
