AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Anthropic says the key to agent memory is knowing what to forget, not how much to store

Collected Nov 16, 2026

Anthropic has published a research post describing how Claude is meant to carry knowledge from one session to the next while the underlying model stays unchanged. The mechanism is deliberately plain: the agent writes notes into a folder and reads them back later. The company's headline finding is that what the notes leave out matters more than how much they hold, and that a follow-up review process can repair notes that turned out to be wrong.

The central difficulty, as Anthropic frames it, is selection. Before the agent knows what happens next, it has to guess which details will matter later, so each note is a bet on the future. Saving everything looks safe but is not: game states and project states move on, so an exhaustive archive fills with outdated content that competes with the current situation for the model's attention.

The Pokémon Red test makes the contrast concrete. The older Claude Sonnet 3.7 behaved as though every observation might be needed, and over roughly 10,000 game steps it produced 413 files spread across 225 folders, frequently a fresh folder for each battle or item it came across. At points it spent its turns reorganizing notes instead of playing, and it finished the run with a single badge. The newer Claude Fable 5 held its memory to a handful of files and collected four badges. It skipped anything already visible on screen and confined itself to plans, strategies and rules it had actually verified in play. Anthropic notes that the comparison covers a single playthrough per model, which is a thin base for a broad claim, though the gap is large enough to be suggestive.

On instructions, Anthropic's recommendation is to explain what the memory is for and leave the method to the model. In one benchmark, rigid rules such as recording only mistakes erased most of the advantage for stronger models. Weaker models moved in the opposite direction: they did better with explicit directions, including a reminder to read their notes before each response. The practical reading is that scaffolding compensates for weaker judgment, while over-specifying constrains a capable model's ability to decide what is worth keeping.

Faulty notes are the cost of the approach. An agent that trusts its own records will act on a mistake repeatedly, which is what happened in the Pokémon run: Claude had written down the wrong side of a mountain for a landmark, fell into the same trap three times, came close to giving up and missed its objective. The fix, which Anthropic calls Dreaming, borrows the sleep metaphor. During idle periods, Claude revisits transcripts of earlier sessions and checks its notes against them, correcting errors and building out the upper layers of its memory into structured collections. After a Dreaming pass, Claude caught the landmark error and found the correct route, reaching Cerulean City past Mt. Moon only because of the correction. The concept builds on prior work from Letta and UC Berkeley.

The same pass is meant to scale to large document sets. Claude assembled a linked wiki from roughly 350,000 internal documents at Stainless, a software company Anthropic acquired, with Dreaming generating new pages and repairing gaps, duplicates and broken links. Answers improved the longer the wiki was worked on, and the best results came from pairing the wiki with search rather than using it as a substitute. In a separate legal test, questions whose answers were scattered across several documents were where the wiki helped most: Claude scored 87 percent with the wiki against 73 percent with search alone. On simple lookups the two were roughly level, and in one case the wiki ended up larger than the source material it summarized.

Anthropic is candid about the open problems. Note-taking during a task pulls the agent away from the work itself. Dreaming consumes extra compute and delivers its corrections only after a delay. Wikis and notes decay without upkeep, so errors and hallucinations accumulate. Loading too many memories into every session is also a trap, since they consume context and any bad note gets inherited by every future session.

Dreaming first appeared in May as a research preview for Claude Managed Agents, Anthropic's cloud platform for agents. It reads an existing memory store plus up to 100 past sessions, strips duplicates and stale entries, and writes a reorganized store while leaving the original untouched; at launch it supported only Claude Opus 4.7 and Sonnet 4.6. The name is now shared: OpenAI also calls ChatGPT's memory feature, overhauled in June, Dreaming, and says the rework lifted temporal accuracy from 52.2 to 75.1 percent. Google and Google DeepMind researchers use dreaming differently, to improve an agent's search strategy by working through alternatives from stored results without another model call, and Google's WikiSkill keeps raw transcripts, distilled knowledge and instructions in separate layers much as Anthropic's system does.

Why it matters: teams running agents over multi-day projects inherit a design decision they often make by default, and Anthropic's results argue that a small, curated, periodically audited memory beats a comprehensive log. The trade-offs are real, since Dreaming costs compute and adds latency, rigid note-taking rules can hurt the strongest models, and a single bad note can propagate into every later session. For product teams, the implication is that memory should be treated as a maintained artifact with review cycles, not an append-only transcript, and that retrieval plus a curated knowledge base may outperform either alone. These are inferences from Anthropic's own tests, which have not yet been independently replicated.

Read at The Decoder

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Anthropic studied how Claude retains knowledge across sessions. Selective notes beat exhaustive archives by a wide margin. In a Pokémon Red test, the newer Claude Fable 5 got four times as far as the older Sonnet 3.7 by keeping just a handful of strategy files instead of more than 400. A "Dreaming" process retroactively corrects faulty memories. The article Anthropic says the key to agent memory is knowing what to forget, not how much to store appeared first on The Decoder .