The easiest way to give an agent memory is to store everything. Anthropic's The components of agent memory, published November 6, 2026, challenges that instinct. What matters, it argues, is what you turn the stored material into, where you keep it, and when you correct it.
Four layers
The post defines agent memory as knowledge Claude retains between sessions. Weight changes, in-context learning and compaction are out of scope.
| Layer | Content | Character |
|---|---|---|
| Transcripts | Searchable event log per session | Raw, foundational |
| Memory notes | Observations Claude writes mid-session | A bet that it will be useful later |
| Memory representations | e.g. a linked wiki on one subject | Notes made into structure |
| Injected context | Loaded into every session | Must stay small |
The offline process that moves between layers is called dreaming: the agent reviews its logs to fix mistakes, write notes and build new representations. Our three layers of agent memory article splits memory by lifetime (context, session, persistent). Anthropic splits it by how processed the stored material is. The two views complement each other.
Evidence 1: the gain depends on the model
Mean reward on the CL-Bench 1.0 database task, as reported by Anthropic:
| Model | No memory | With memory |
|---|---|---|
| Opus 5.5 | 0.655 | 0.838 |
| Sonnet 5.5 | 0.644 | 0.833 |
| Opus 5 | 0.564 | 0.828 |
| Opus 4.7 | 0.383 | 0.700 |
| Sonnet 5 | 0.306 | 0.610 |
| Sonnet 4.6 | 0.330 | 0.364 |
Sonnet 4.6 barely moved. The post describes it as generating theories without testing them. Prompting it to read memory and test theories lifted it from 0.364 to 0.641. Narrow instructions can hurt: telling models to record only failures cost Opus 5 0.234. The authors' takeaway is that for stronger models, explaining what memory is for beats prescribing fixed rules.
Evidence 2: wikis work best on top of search
On a company corpus of about 350,000 files (about 420M tokens) with 25 questions:
- Wiki alone went from 13% to 48% across 2 to 128 sessions.
- Wiki plus search went from 54% to 61%; search alone scored 53%.
- Wiki plus search saved about 20,000 tokens per question.
The wiki improved with more sessions but did not replace search. On some tasks the gain was unclear because one text search was enough. In another case the wiki grew to 1.77 times the 115 KB corpus; the post compares it to a map as large as the empire it describes.
Evidence 3: wrong memory spreads, and can be repaired
In the Pokémon Red experiment, a wrong note put the PokéCenter on the wrong side of a mountain. After dreaming corrected it, the agent reached its target city at action 1,935; the uncorrected run and the no-memory run never did. In a contract-brief task, agents writing to one shared table kept 41 of 57 facts (72%). Giving each contract its own page kept all 57. That is the basis for the one-writer-per-document advice.
Applying it to your own agent
Practical points drawn from the post:
- Make transcript search the base layer and add a wiki on top.
- Keep injected context short. The post says Claude Code reads only the first 200 lines or 25 KB of
MEMORY.md, because one wrong line is carried into every session. - Store what the screen cannot show (strategy, plans), not state you can re-read each turn.
- Record source and verification status in notes, and recheck volatile facts first. This last item is my generalization of the post's dreaming advice, not a claim from it.
Limits
- The benchmark numbers are Anthropic's own runs; I found no independent reproduction.
- Dreaming costs compute, and the post concedes that continuously checking all memory may be impractical.
- Whether to write memory in the moment or during later consolidation is left open.
Conclusion
The data does not support "store more and it gets better." Memory helps by very different amounts depending on the model and task, and a wrong memory scales its damage. Design effort belongs in writer separation, a short injected layer, and regular review rather than capacity. For a contrasting approach, see the Memorilla paper's latent memory tokens, where memory a human cannot read creates its own audit problem.
Sources
- Anthropic, The components of agent memory (2026-11-06, checked 2026-11-07)
Operations note: keep provenance with the memory
The primary source was reviewed again on November 10, 2026. This update adds operational guidance rather than publishing the same study as a new announcement. The following is editorial advice, not a measured result from the study.
For each note, retain a source location, observation time, verification state, and the identifier of any note it supersedes. Set expiry according to how quickly the fact changes: a price or permission can age differently from a long-term strategy. Expiry can trigger source review rather than immediate deletion.
Prioritize recently used and change-prone notes for revalidation, and preserve the old text and correction reason. Against a search-only baseline, measure answer accuracy, wiki maintenance tokens, revalidation delay, and the rate of answers citing stale memories. The company’s 25-question evaluation over 350,000 files does not establish the effectiveness of this proposed operational policy.