Anthropic publishes its working theory of agent memory as Claude learns to dream between sessions

Anthropic publishes its agent memory architecture

Anthropic has published its working theory of how Claude should remember what it learns between sessions. In a detailed research post published on November 6, 2026 by Lance Martin of Anthropic Engineering, the company lays out three components of agent memory: in-session notes, a learned sense of salience, and dreaming — an offline pass where Claude reviews its own session transcripts to correct mistakes and reorganize knowledge.

The post, titled The components of agent memory, draws an explicit analogy to human memory. The hippocampus captures experiences in the moment, but only some are consolidated into long-term knowledge during sleep and quiet rest. Anthropic's design mirrors that division: Claude writes memory notes during a session, and dreaming replays and reorganizes them between sessions.

Agent memory here means knowledge Claude gains from experience — interactions with users and tasks it has been assigned — that it retains between sessions as saved context it can read later. It excludes changes to model weights, in-context learning within a single conversation, and compaction (summaries that replace older turns within a context window).

How Claude writes memory notes and judges what matters

Claude can tag salient events in a session log as they happen and write them to a memory store — a folder of plain text files it can read in future sessions. The key question is how Claude decides what is salient before it knows what will matter later.

Anthropic frames memory as a prediction about the future: by writing down a rule, Claude predicts the pattern will be relevant again. Writing down a verified fact means predicting a future session will want to rely on it without double-checking. Over model generations, Claude has become much better at making these predictions.

The company tested this on Continual Learning Bench (CL-Bench) 1.0, a benchmark that measures an agent's ability to use memory for long-horizon tasks. Claude had to answer questions by looking up a database, then after each answer it was told the correct solution and could write a memory to help with future questions.

Different models handled this differently. Claude Sonnet 4.6 generated theories about why an answer was wrong, but its memory directory held scattered, untested guesses. Opus 4.7 consolidated notes into a single reference document and marked which facts were checked and which unconfirmed. Sonnet 5 marked each answer as confirmed after grading and wrote corrections when it had answered wrong.

Opus 5 was the first to close the loop: most entries included a rule, how the rule was confirmed, answers to earlier questions, and clear warnings about traps. Opus 5.5 and Sonnet 5.5 both closed the loop as well, scoring similarly to Opus 5 with memory active.

Pokémon Red reveals what strong models remember differently

Anthropic used Pokémon Red as an unusual test of memory judgment. At the start of each run, Claude got a folder of memory files it could read from and write to while playing. The results showed a stark difference in what each model thought was worth remembering.

Claude Sonnet 3.7 wrote 413 memory files across 225 folders and earned one badge (badges mark progress through the game). It assumed everything might matter. Its archive held separate files for each battle, each item found, and each non-player character met. Around step 10,000, Sonnet 3.7 started a fourth folder of "master" memory files. Over the next 17 steps, it created six files, deleted nine others, and took no actions in the game itself.

Claude Fable 5 wrote nine files in two folders, plus one top-level file called global.md, and earned four badges. Its memory was organized strategically: global.md held the current goal, a plan for tackling gyms, and the model's own rules for using memory. A locations folder held route navigation plans, and a gameplay folder held strategies based on what it had learned.

The crucial difference was that Fable 5 stored only what the game did not explicitly tell it — strategy, plans, and tested rules — while Sonnet 3.7 stored information it could read off the screen every turn, like money or map position. Opus 5.5 followed the same pattern: a small folder of notes tracking its current goal and next steps, rewritten as it played.

Steering memory: giving models freedom beats fixed rules

Anthropic tested how much guidance Claude needs to use memory well. Simply telling Claude a memory folder was available was enough for Opus 4.7 and Opus 5 to get a noticeable performance boost. Giving stronger models fixed rules about what to write — like "record only failures" — actually reduced their performance.

"For stronger models, giving them freedom to judge salience was more effective than setting fixed rules in advance, and those rules could end up weakening performance," Anthropic writes.

Steering helped older models the most. Prompting Sonnet 4.6 to read its memory files before answering and to explicitly test its theories nearly doubled its benchmark score. Sonnet 5.5 did this on its own — reading memory before answering 9 out of 10 times — while Sonnet 4.6 did it less than half the time without being prompted.

Dreaming: how Claude corrects its own memory offline

The most novel component Anthropic describes is dreaming. Between sessions, Claude can review its own session transcripts — the raw logs of what happened — and check whether its memory notes are correct. If a note was wrong from the start or has drifted as the environment changed, dreaming lets Claude catch and fix it.

Anthropic demonstrated this with the Pokémon Red game. At one point Claude wrote down that a PokéCenter was on the east side of Mt. Moon when it was actually on the west. This wrong note sent Claude over a one-way ledge repeatedly, trapping it in a loop. After the third failure, the model seemed to give up, writing: "I'm going to stop rather than keep looping."

When Anthropic ran dreaming on the memory folder containing the wrong location, Claude checked the note against the session transcript, caught its error, and rewrote it. With the corrected note, Claude went straight through Mt. Moon and reached the next city. Without dreaming, it had fallen off the ledge three times and never made it through.

Anthropic describes the dreaming pass mechanism as a self-verification step: Claude checks its memory notes against the raw transcript of what happened, catches discrepancies, and updates the memory store before the next session starts. That let Claude compare its notes against what actually happened in the session and fix mistakes.

Dreaming can build wiki-style knowledge bases

Beyond correcting errors, dreaming lets Claude build new memory representations. In one evaluation, Anthropic gave Claude access to about 350,000 pages of documentation for Stainless, a functional programming library. Over a series of dreaming passes, Claude read documentation pages, wrote individual notes, and gradually organized them into a linked wiki — cross-referenced entries with navigation, indexes, and summaries.

The resulting wiki was not a flat dump. Claude grouped related concepts, created an index page, and linked entries to each other. A future session that needed to work with Stainless could load the wiki and navigate to relevant facts without reading and re-reading all 350,000 pages.

This goes beyond summarization. The dreaming process involved Claude reading pages, writing individual memory notes, then reading those notes in later dreaming sessions to merge, consolidate, and link them — an iterative knowledge-building loop that mirrors how a human might gradually build expertise in a codebase.

What the findings mean for agent developers

Anthropic's post carries practical implications for anyone building agents on Claude. The company now recommends simple guidance to create memory notes when a person asks or when something is clearly worth remembering, and leaving the judgment of what to keep up to Claude. Fixed rules about what constitutes important memory appear to hurt stronger models.

The effectiveness gap between models is also notable. Teams running Sonnet 4.6 or earlier models may need explicit prompting to use memory at all, while Opus 5 and newer models handle salience judgment autonomously. Anthropic says that with frontier Claude models, it typically gives simple guidance to create memory notes and leaves the judgment up to Claude.

The dreaming process — offline review of session transcripts — is available through Claude Managed Agents with a mounted memory directory, which Anthropic opened for dynamic workflows on October 16. The company has not announced general availability of dreaming as a standalone feature, but the research post indicates it is already part of how Anthropic builds Claude memory internally.

The limits of Anthropic's approach

Anthropic is transparent about what the post does not cover. The evaluations are Anthropic's own and were not reproduced by an independent lab. CL-Bench scores compare different models on different evaluation runs. The Pokémon Red test is a qualitative comparison, not a rigorous benchmark — Sonnet 3.7 has been available for over a year while Fable 5 shipped recently.

The post also does not address how dreaming scales. Reviewing session logs means storing transcripts, and the compute cost of an offline consolidation pass increases with the number of sessions and size of the memory store. For agents that run for months or handle sensitive data, the storage, privacy and cost implications of keeping full session logs for dreaming are not discussed. And the approach is specific to Claude — other models would need their own infrastructure to implement similar memory systems.