Key takeaway
Anthropic published its agent memory research on November 6, 2026: Claude writes reusable notes during sessions, then an offline "dreaming" phase checks notes against transcripts, corrects errors, and builds wikis. On an internal corpus of about 350,000 files and 420 million tokens, a 25-question score rose from 13% to 48% as dreaming scaled, and a wiki plus search reached 61% versus 53% for search alone.
What happened
On November 6, 2026, Anthropic's engineering team published "The components of agent memory" on its research blog. The post defines agent memory as knowledge Claude gains from experience and retains between sessions as context it can read later, and explicitly separates it from weight changes, in-context learning within a single conversation, and compaction.
The method splits memory into two stages. During a session, Claude tags salient events and writes memory notes. Between sessions, a phase called dreaming replays logs, validates notes against the original transcripts, fixes mistakes, and produces new memory representations. Anthropic credits Letta and UC Berkeley's "sleep-time compute" idea, which uses idle time between interactions to reorganize context.
Key details
| Item | What Anthropic published |
|---|---|
| Publish date | November 6, 2026 |
| Memory format | Persistent files (notes, folders, wiki pages); model weights are not changed |
| Role of dreaming | Replay session logs, validate notes, correct errors, build wikis and other retrievable representations |
| Evaluation corpus | About 350,000 files and roughly 420 million tokens from Stainless, acquired in May 2026 |
| Evaluation scale | 25 questions, 128 dreaming sessions, Opus 5 agents |
| Wiki-only answering | 13% with two sessions rising to 48% with 128 sessions, measured against graded criteria |
| Wiki plus search | Aggregate score rose from 54% to 61%; search alone scored 53% |
| Token cost | Wiki plus search saved about 20,000 tokens per question |
| Legal benchmark | Harvey LAB merger task, 24 questions: 87% with the wiki versus 73% with search alone |
Two layers: notes and dreaming
Anthropic frames writing memory as a prediction about future need: recording a verified fact predicts that a later session can rely on it without double-checking. Success at this varies by model generation. On a Continual Learning Bench 1.0 database task, Sonnet 4.6 produced scattered files of untested guesses; Opus 4.7 consolidated notes into one reference document and marked which facts were checked; Sonnet 5 marked answers as confirmed only after the grader verified them; Opus 5 was the first to close the loop, with most entries containing a rule, how the rule was verified, and warnings about traps. Opus 5.5 and Sonnet 5.5 also closed the loop.
The Pokémon Red evaluation quantified how much the choice of what to save matters. About 10,000 steps into the game, Sonnet 3.7 had written 413 memory files across 225 folders and earned one badge, while Fable 5 wrote nine files in two folders plus a top-level global.md and earned four badges. Anthropic's analysis: Sonnet 3.7 stored state that is visible on screen every turn, such as money and position, so its records went stale and it spent more steps curating memory than playing. Fable 5 stored only strategy, plans, and tested rules that the game does not state.
Dreaming gives notes a ground truth. In the Mt. Moon example, Claude incorrectly recorded the PokéCenter as being on the east side, announced "Mt Moon DONE" after leaving it, and fell back down a one-way ledge three times before writing that it would stop. Dreaming compared the note against the session transcript, caught the reversed location, and rewrote it; with the corrected note, Claude re-entered the cave and reached the next city.
Results and cost
The largest evaluation used Stainless's internal documents, about 350,000 files and roughly 420 million tokens. Anthropic built 25 questions and had Opus 5 agents build a wiki over 128 sessions, exploring files, writing and correcting linked pages, merging duplicates, and checking claims against sources.
According to the post, scaling dreaming from two sessions to 128 raised performance on the 25 questions from 13% to 48% of graded criteria. When the wiki was combined with a search tool, the aggregate score rose from 54% to 61% as the wiki grew, compared with 53% with search alone, while saving about 20,000 tokens per question.
In the legal setting, the Harvey LAB merger task included 24 questions, more than half requiring information from several documents; Claude scored 87% when consulting the wiki versus 73% with search alone. A separate contract task exposed a concurrency problem: several agents wrote to one shared table of contract terms, and only 41 of 57 facts (72%) survived. After each contract got its own page written by a single agent quoting terms verbatim, all 57 facts survived and 100% of graded criteria could be met.
Limits and scope
Anthropic lists several constraints that rule out applying the method unconditionally:
- Wrong notes get reinjected. A memory that is wrong from the start, or preloaded into every session, propagates bad decisions.
- Wikis can outgrow their sources. In one LAB task, Claude's wiki became larger than the documents it summarized, abstracting nothing.
- Retrieval remains necessary. When an answer sits in a single file, direct search performs about as well; the wiki's advantage appears on tasks that require combining documents.
- Cost lands outside the session. Dreaming consumes compute, so it pays off only when tasks are long and repetitive enough to amortize consolidation.
- The injected layer must stay small. Claude Code's auto memory loads only the first 200 lines or 25 KB of MEMORY.md as an index, and Anthropic's documentation warns that longer instruction files reduce adherence.
Anthropic also reports that for stronger models, simple guidance about what memory is for works better than fixed rules. When it tested a "record only failures" rule, both Opus 5 and Opus 4.7 lost much of the gain that memory otherwise provided.
Industry impact
Three transferable conclusions follow. First, agent memory is less about a larger context window than about judging what to write and when to validate it. Second, offline consolidation separates memory upkeep from the main task, at the price of extra compute and latency. Third, memory systems should be measured on task completion combined with token cost, not on recall alone.
Facts and assessments should be separated: the scores and cases above come from Anthropic's own post, so they are vendor-reported, and the post itself states that the test scale was limited and the results do not generalize to all settings. Whether to adopt the approach depends on whether a workload actually involves repeated cross-session, cross-document queries.
Official sources
- Anthropic: The components of agent memory (November 6, 2026)
- Anthropic Cookbook: Dreaming with a dynamic workflow