Anthropic lays out agent memory as transcripts, notes, and 'dreaming' consolidation
- Anthropic describes agent memory as four layers: transcripts (a searchable event log), memory notes Claude writes, memory representations such as wikis, and context injected into every session, with dreaming moving information up from transcripts.
- Dreaming builds on sleep-time compute, proposed by Letta and UC Berkeley researchers: a model uses idle time between interactions to reorganize its context, here by reflecting on logs to fix mistakes and write new notes.
- Anthropic's definition excludes weight updates, in-context learning, and compaction; agent memory means knowledge gained from users and tasks that persists between sessions as context Claude can read later.
- On a CL-Bench 1.0 database lookup task, Claude starts each question in a fresh session and, after being told the correct answer, can record the failure, test theories about it, and turn the result into a rule for later questions.
- According to figures quoted in the discussion, prompting Claude to read its memory files and explicitly test its theories nearly doubled its benchmark score, with Opus 5 gaining 46.8% and Opus 4.7 gaining 82.7%.
Hacker News 의견들
Read half of it and had to stop. I prototyped a lot of this about 2.5 years ago and never shipped it, and seeing it laid out this cleanly stings. Fig. 3 is the part that got me: Opus 5.5, Sonnet 5.5 and Opus 5 all tied on the basic memory test.
The 'I had the idea first and did nothing with it' feeling is the worst. Only thing that's helped me is rolling more dice: share everything, find people who get excited about it. The labs just have infinite dice.
I don't see anything seminal here. It's markdown files plus a prompt telling the model to reread them, dressed up with a neuroscience analogy.
Most lab posts are easy jumps from current harness behavior, honestly. There are probably n open source projects doing roughly this. What's actually interesting is the empirical evidence, and that costs millions of dollars to produce.
This one is mostly marketing. Calling 'write a note to an .md file and reload it later' dreaming is a stretch, and Claude rarely actually remembers anything. Mine keeps apologizing and writing a 15000th note with the same content.
Dreaming is an analogy for post-hoc conflict resolution in memory, which is one of the real things dragging down long-run performance. Doesn't read as a consciousness claim to me.
It's unscientific to the point of superstition. I wouldn't be shocked if Anthropic reincorporated as a religion.
Disagree on the useless experiment bit. Comparing memory folders from Sonnet 3.7 and Fable 5 at 10,000 steps is testing exactly the failure you just described. And 1M tokens isn't a tiny context.
I turned off automatic memory creation in every harness I use. Claude Code still writes some and I delete about 95% of them, because they only matter for the task in front of it. Better judgment about what to save is the improvement I want.
The dreaming section made me realize transcripts are underused. Dump your own into duckdb, Claude Code will build the pipeline in ten minutes, then look at tool call failures and cache misses and tokens per dollar for main thread vs subagents.