<p>Sounds like a joke, but it’s not. The Claude Code community is discussing a new feature — Auto Dream. The agent literally “falls asleep” and tidies up its memory.</p>
<p>Disclaimer: for now, it resembles more of a research feature. Some users report that it doesn’t work for them. There hasn’t been an official announcement from Anthropic. However, the system prompt has already been found on GitHub, and the mechanics have been detailed.</p>
<h2>What it solves</h2>
<p>A couple of months ago, Claude Code received Auto Memory — the agent records notes about the project between sessions: build commands, architectural decisions, code style. But by the 20th session, the memory turns into a mess: outdated records, contradictions, noise. The agent starts to perform worse.</p>
<p>Auto Dream is a garbage collector for AI memory. It works in 4 phases:</p>
<ul>
<li>→ Orient — scans what has already been saved</li>
<li>→ Gather signal — compares memory with the actual state of the code</li>
<li>→ Consolidate — merges, removes duplicates, replaces “yesterday” with specific dates</li>
<li>→ Prune — cleans dead links and contradictions</li>
</ul>
<p>It runs in the background, only after 24 hours + 5 sessions. Read-only on the code, it only writes to memory files.</p>
<h2>Why it’s interesting</h2>
<p>The analogy with REM sleep is not just marketing. During rapid eye movement sleep, the brain does exactly the same: it processes daily experiences, discards garbage, strengthens the important, resolves contradictions. We are literally modeling AI after neurobiology.</p>
<p>Skeptics rightly point out: if the LLM decides what is important and what is not — it may delete something critical. And there are currently no public evaluations confirming real improvement.</p>
<p>Source: <a href="https://www.youtube.com/watch?v=OnQ4BGN8B-s">https://www.youtube.com/watch?v=OnQ4BGN8B-s</a></p>;
<p>#ai #claude #dev</p>
<p><a href="https://t.me/+OvImEUmA7W5mYTRi">————————— Мысли Рвачева —————————</a></p>
