<?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: Sergei Gordeichuk</title>
    <description>The latest articles on DEV Community by Sergei Gordeichuk (@gosen).</description>
    <link>https://dev.to/gosen</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%2F3763893%2F500c3183-600b-4dbe-b9a4-a7e7e9499671.jpg</url>
      <title>DEV Community: Sergei Gordeichuk</title>
      <link>https://dev.to/gosen</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/gosen"/>
    <language>en</language>
    <item>
      <title>There is no agent memory format standard. I counted what my agent remembers instead</title>
      <dc:creator>Sergei Gordeichuk</dc:creator>
      <pubDate>Mon, 16 Nov 2026 11:56:40 +0000</pubDate>
      <link>https://dev.to/gosen/there-is-no-agent-memory-format-standard-i-counted-what-my-agent-remembers-instead-4b38</link>
      <guid>https://dev.to/gosen/there-is-no-agent-memory-format-standard-i-counted-what-my-agent-remembers-instead-4b38</guid>
      <description>&lt;p&gt;There's no agent memory format standard as of 16 November 2026. I found no vendor that ships a coding agent in any memory spec effort, the &lt;a href="https://aaif.io/projects/" rel="noopener noreferrer"&gt;Agentic AI Foundation&lt;/a&gt; lists ten projects and none of them is about memory, and the one standards body group on the subject (a &lt;a href="https://www.w3.org/community/ai-agent-memory-interop/" rel="noopener noreferrer"&gt;W3C Community Group&lt;/a&gt; chartered on 19 June 2026) hasn't published a spec. The open proposals are papers and small repositories, and the three with a repository of their own have 2, 7 and 40 GitHub stars.&lt;/p&gt;

&lt;p&gt;The question came up twice in one week, Anthropic published &lt;a href="https://www.anthropic.com/research/components-of-agent-memory" rel="noopener noreferrer"&gt;The components of agent memory&lt;/a&gt; on 6 November and Mads Thines of LoreKit published &lt;a href="https://www.lorekit.io/blog/open-agent-memory-format" rel="noopener noreferrer"&gt;Agent memory needs its OpenTelemetry moment&lt;/a&gt; the next day.&lt;/p&gt;

&lt;p&gt;So I'd stop waiting for a standard, and I wouldn't adopt a memory store to get portability either. Put whatever has to survive a change of harness into AGENTS.md and Agent Skills, and treat each tool's own memory as a cache you can lose. I checked what that cache holds on my own machine.&lt;/p&gt;

&lt;h2&gt;
  
  
  113 memory records, none of them portable
&lt;/h2&gt;

&lt;p&gt;Claude Code auto memory on my development Mac holds 113 records in 33 project directories (counted with Claude Code 2.1.323). I read the front matter keys of each file and nothing else, and here's what the script printed.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight console"&gt;&lt;code&gt;memory records: 113
front matter keys &lt;span class="o"&gt;(&lt;/span&gt;records carrying each&lt;span class="o"&gt;)&lt;/span&gt;:
  description: 113
  type: 113
  name: 113
  originSessionId: 105
  metadata: 97
  node_type: 94
  modified: 54
fields a transfer would need &lt;span class="o"&gt;(&lt;/span&gt;records carrying each&lt;span class="o"&gt;)&lt;/span&gt;:
  expiry field: 0
  &lt;span class="nb"&gt;source&lt;/span&gt; or citation field: 0
  confidence field: 0
  writer field &lt;span class="o"&gt;(&lt;/span&gt;agent, model or tool&lt;span class="o"&gt;)&lt;/span&gt;: 0
  body links another memory with &lt;span class="o"&gt;[&lt;/span&gt;&lt;span class="o"&gt;[&lt;/span&gt;name&lt;span class="o"&gt;]&lt;/span&gt;&lt;span class="o"&gt;]&lt;/span&gt;: 68
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Three things in that output matter if you ever want to move these records.&lt;/p&gt;

&lt;p&gt;The only provenance goes stale. &lt;code&gt;originSessionId&lt;/code&gt; names a session transcript, but Claude Code deletes transcripts after &lt;code&gt;cleanupPeriodDays&lt;/code&gt; and the memory files stay.&lt;/p&gt;

&lt;p&gt;There are two layouts and no version field, sixteen records have a flat &lt;code&gt;type&lt;/code&gt; key and 97 nest it under &lt;code&gt;metadata&lt;/code&gt;. One tool wrote both within nine months, so a converter has to guess which one it's reading.&lt;/p&gt;

&lt;p&gt;And the links are local, 68 records point at another memory with &lt;code&gt;[[name]]&lt;/code&gt;, which resolves only inside that directory.&lt;/p&gt;

&lt;p&gt;The types say what I'd lose. Of the 113 records, 57 are &lt;code&gt;project&lt;/code&gt; state, 40 are &lt;code&gt;feedback&lt;/code&gt; and 16 are &lt;code&gt;reference&lt;/code&gt;. A &lt;code&gt;feedback&lt;/code&gt; record is a correction I gave the agent, so that's 40 standing instructions that live outside the repo, on one machine, where only one tool looks.&lt;/p&gt;

&lt;p&gt;And my own portability mechanism turned out to be a symlink 😄&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight console"&gt;&lt;code&gt;$ ls -l ~/.codex/AGENTS.md
lrwxr-xr-x  20  ~/.codex/AGENTS.md -&lt;span class="o"&gt;&amp;gt;&lt;/span&gt; ../.claude/CLAUDE.md
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Codex and Claude Code read the same global instructions because I pointed one file at the other, and nothing does that for the 113 records.&lt;/p&gt;

&lt;h2&gt;
  
  
  What already ports, and what does not
&lt;/h2&gt;

&lt;p&gt;Three layers of the agent stack have a shared format today, and memory isn't one of them.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Layer&lt;/th&gt;
&lt;th&gt;Shared format&lt;/th&gt;
&lt;th&gt;Who writes it&lt;/th&gt;
&lt;th&gt;What a switch costs&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Instructions&lt;/td&gt;
&lt;td&gt;AGENTS.md&lt;/td&gt;
&lt;td&gt;A person&lt;/td&gt;
&lt;td&gt;Nothing, a symlink or one setting&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Procedures&lt;/td&gt;
&lt;td&gt;Agent Skills (&lt;code&gt;SKILL.md&lt;/code&gt;)&lt;/td&gt;
&lt;td&gt;A person, sometimes an agent&lt;/td&gt;
&lt;td&gt;Nothing where the harness reads skills&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Tool calls&lt;/td&gt;
&lt;td&gt;Model Context Protocol&lt;/td&gt;
&lt;td&gt;A server author&lt;/td&gt;
&lt;td&gt;Re-registering the server&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Enforcement&lt;/td&gt;
&lt;td&gt;None, hooks differ per harness&lt;/td&gt;
&lt;td&gt;A person&lt;/td&gt;
&lt;td&gt;Rewriting the hooks&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Memory&lt;/td&gt;
&lt;td&gt;None&lt;/td&gt;
&lt;td&gt;An agent or a service&lt;/td&gt;
&lt;td&gt;The memory&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;MCP is often named as the answer for memory, but it covers the transport only. In Thines's words, "MCP standardises the call, not the record." A memory store reached over MCP moves with you from harness to harness and then ties you to that store.&lt;/p&gt;

&lt;p&gt;GitHub Copilot is the other end. Its memory is the best-specified shape I found, facts carry citations, the agent and model that wrote them and an expiry date (90 days by default). It's also the least portable, the &lt;a href="https://docs.github.com/en/copilot/concepts/prompting-and-context/copilot-memory" rel="noopener noreferrer"&gt;docs&lt;/a&gt; say "The SDK does not provide APIs to list, edit, or delete memories."&lt;/p&gt;

&lt;h2&gt;
  
  
  Why I would not wait
&lt;/h2&gt;

&lt;p&gt;The OpenTelemetry comparison argues for patience more than for a spec. OpenTracing and OpenCensus &lt;a href="https://medium.com/opentracing/merging-opentracing-and-opencensus-f0fe9c7ca6f0" rel="noopener noreferrer"&gt;announced their merger&lt;/a&gt; on 28 March 2019 and the stable tracing specification came on 10 February 2021, about 22 months later, and that clock started with two projects that already had users and corporate backers. Agent memory doesn't have such a pair. If two major vendors announced a joint format tomorrow, the same pace would put a stable record in the second half of 2028.&lt;/p&gt;

&lt;p&gt;The products are also still moving. Gemini CLI's current docs have no &lt;code&gt;save_memory&lt;/code&gt; tool, Copilot's default expiry went from 28 days to 90, and Codex stores memories in a directory named &lt;code&gt;memories_v2&lt;/code&gt; with an undocumented file format.&lt;/p&gt;

&lt;h2&gt;
  
  
  The best case against this
&lt;/h2&gt;

&lt;p&gt;People don't write things down, and agents do. An agent records the correction at the moment it's made, and of my 113 records 79 include a line explaining why the note exists, which is more than I would've written myself 😄 Long instruction files are also followed less well, so moving memory into AGENTS.md has a price.&lt;/p&gt;

&lt;p&gt;I accept both points, but they're reasons to keep agent memory switched on and neither makes it portable. A captured note is durable only once it sits where the next agent loads it.&lt;/p&gt;

&lt;p&gt;I'd be wrong if one of Anthropic, GitHub, OpenAI, Cursor or Google joined a memory spec effort, if the Agentic AI Foundation opened a memory project, or if two shipping harnesses began loading each other's memory directory. None of that had happened by 16 November 2026.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I would do
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;If you run two or more coding agents, make AGENTS.md the real instruction file and point the others at it with a symlink or an import.&lt;/li&gt;
&lt;li&gt;If your agent writes memory, read the directory once a month, move each standing correction into AGENTS.md or a skill (in a commit a person reviews) and delete the note.&lt;/li&gt;
&lt;li&gt;If a note is project state, such as a decision or a deadline, let it expire with the tool, it belongs in the issue tracker.&lt;/li&gt;
&lt;li&gt;If you use Copilot Memory, treat its repository facts as Copilot's working notes. Anything the team relies on goes into the repo.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The full post has the seven spec proposals with their status, the five memory shapes side by side and what each vendor changed in the last twelve months: &lt;a href="https://stackness.dev/blog/agent-memory-format-standard-none-exists-what-ports-between-agents" rel="noopener noreferrer"&gt;Agent memory format standard: none exists, so keep what must port in the repo&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Have you opened your agent's memory directory? I'm curious what other people found in theirs.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>agents</category>
      <category>claudecode</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Kantamigo for your guitar parties</title>
      <dc:creator>Sergei Gordeichuk</dc:creator>
      <pubDate>Sat, 07 Nov 2026 17:44:19 +0000</pubDate>
      <link>https://dev.to/gosen/kantamigo-for-your-guitar-parties-5648</link>
      <guid>https://dev.to/gosen/kantamigo-for-your-guitar-parties-5648</guid>
      <description>&lt;p&gt;Syncing a lyrics scroll across 30 phones without sending pixels&lt;/p&gt;

&lt;p&gt;I'm building Kantamigo, a web app for guitar nights where friends follow the lyrics on their phones while the guitarist plays. The fun problem was the scroll because the guitarist's screen and every guest's phone have different widths, fonts and line wraps, so a pixel offset means nothing on the other side.&lt;/p&gt;

&lt;p&gt;So positions are line based (for now). The performer sends {item, section, line, frac, playing, speed}, where frac is how far into the current line we are and each phone works out its own pixels from that. Lines wrap differently on a small phone, but "verse 2, line 3, halfway" is the same everywhere&lt;/p&gt;

&lt;p&gt;The performer sends at most 10 updates a second plus a heartbeat every 2 seconds, and between updates the phones keep scrolling at the given speed so it looks smooth without a flood of messages.&lt;/p&gt;

&lt;p&gt;The API runs on more than one instance, and a guest's socket can sit on a different one than the guitarist's. Every instance subscribes to a Redis pub/sub channel per session and forwards events to its local sockets.&lt;/p&gt;

&lt;p&gt;Phones drop off all the time at a party (locked screens, bad wifi in the kitchen), so a client never relies on events it missed. On reconnect it gets a full state snapshot and rebuilds from that.&lt;/p&gt;

&lt;p&gt;Backend is Go with coder/websocket, frontend is Next.js. Feel free to check it out and please please let me know what do you think: &lt;a href="https://kantamigo.com" rel="noopener noreferrer"&gt;https://kantamigo.com&lt;/a&gt;&lt;/p&gt;

</description>
      <category>showdev</category>
      <category>go</category>
      <category>websockets</category>
      <category>nextjs</category>
    </item>
    <item>
      <title>Stackness launch</title>
      <dc:creator>Sergei Gordeichuk</dc:creator>
      <pubDate>Sat, 12 Sep 2026 15:01:11 +0000</pubDate>
      <link>https://dev.to/gosen/stackness-launch-54fo</link>
      <guid>https://dev.to/gosen/stackness-launch-54fo</guid>
      <description>&lt;p&gt;Hey everyone,&lt;/p&gt;

&lt;p&gt;Recently I've launched Stackness - a social platform for developers to show their stack and how they use it, and I'd be very grateful to get some feedback from the community! I've launched it on Product Hunt today as well so first users are slowly arriving :)&lt;/p&gt;

&lt;p&gt;Please take a look on my profile there: &lt;a href="https://stackness.dev/stck/gordeychuk_s" rel="noopener noreferrer"&gt;https://stackness.dev/stck/gordeychuk_s&lt;/a&gt;, and let me know what you like and don't like about Stackness in general, thanks a lot! &lt;/p&gt;

</description>
      <category>webdev</category>
      <category>productivity</category>
      <category>showdev</category>
    </item>
    <item>
      <title>Do you get dopamine kicks when Claude Code / Codex does a good job?</title>
      <dc:creator>Sergei Gordeichuk</dc:creator>
      <pubDate>Tue, 17 Feb 2026 13:16:48 +0000</pubDate>
      <link>https://dev.to/gosen/do-you-get-dopamine-kicks-when-claude-code-codex-does-good-job-2a23</link>
      <guid>https://dev.to/gosen/do-you-get-dopamine-kicks-when-claude-code-codex-does-good-job-2a23</guid>
      <description>&lt;p&gt;I actually do, and it happens quite often. I can even feel addicted to it sometimes.&lt;/p&gt;

&lt;p&gt;A little bit about myself: I'm engineering software for ~16 years, and I LOVE it. I literally love everything about it - architecture, plan, coding itself, deployment, debugging, testing, talking to users, building interfaces, all of it. I mostly work with web, but I worked in different areas - TV broadcasting, ads, Wi-Fi chipsets, fitness machines, banking, etc. So I know what it feels like to get a dopamine kick when I "finally resolved that bug!" which I've been working on for days. I've been having that feeling since age 14 when I wrote my first code in QBasic.&lt;/p&gt;

&lt;p&gt;And I feel like these days I can get the same fun, but faster. Because my dev environment with agents, virtualisation, and cool workflows helps me to architect and engineer rather then just code. That means I get dopamine shots way more often. It's like tiktok (which I hate) over long youtube videos. Of course, I iterate, make things better and better, but this initial dose, when the AI agent implemented things exactly the way I wanted it - it's amazing. &lt;/p&gt;

&lt;p&gt;Do you get those? Shall we already gather a group of anonymous ClaudeCode-addicted people? &lt;/p&gt;

&lt;p&gt;An "obvious" disclaimer: any analogy to drugs is to explain my idea, drugs are awful and everybody should avoid them.&lt;/p&gt;

</description>
      <category>programming</category>
      <category>ai</category>
      <category>webdev</category>
      <category>discuss</category>
    </item>
  </channel>
</rss>
