Synergised Consulting
Technical asset

Does more memory make an AI agent worse? Anthropic's research says it depends what you store

Synergised Consulting Ltd

7 min read
Title card: the words Does more memory make an AI agent worse? on a pale ground with an orange rule, above the caption First 100 Days Operations

Does more memory make an AI agent worse? It depends what the memory holds. Anthropic's engineering research, published on 6 November 2026, found that an agent which stores everything it encounters goes backwards, while one that stores only judgement, plans and tested rules progresses further with a fraction of the files. The difference was not the amount remembered. It was what the agent thought was worth remembering.

Last Updated: 7 November 2026

What Anthropic actually published

Anthropic's engineering team has set out how it designs and evaluates persistent memory for Claude agents [1]. The research defines agent memory as knowledge gained from experience that is retained between sessions as context the agent can read later. It deliberately excludes three other things: changes to the model's weights, in-context learning within a single conversation, and compaction, where summaries replace older turns so work can continue in a fresh context window. That exclusion matters for anyone shopping for "memory" features, because a product that only resummarises conversations is not doing this.

The structure Anthropic describes is a hierarchy. At the bottom sit transcripts, the searchable log of everything the agent did. Above them, memory notes: short observations the agent wrote down because it judged them salient in the moment. Above those, representations such as wikis, pages the agent writes and maintains that link related knowledge together. At the top, a small amount of injected context that loads into every session. Only the top layer is free; everything else costs a search.

The research also describes "dreaming", a background consolidation pass in which the agent re-reads its own transcripts and reorganises its memory with hindsight: correcting notes that turned out wrong, building wiki pages, promoting a few notes into the injected layer, and pruning the rest back down. The name borrows from sleep research, where the brain replays recent experience offline and distils it into long-term knowledge.

Two agents, one game, very different notebooks

The clearest evidence in the research is a comparison of two Claude models playing Pokémon Red, each with a memory folder it could read and write during play. According to Anthropic's research [1], at roughly 10,000 steps into the game, Claude Sonnet 3.7 had written 413 memory files across 225 folders and had earned one badge. Claude Fable 5, playing the same game with the same tools, had written nine files in two folders plus a single top-level file, and had earned four badges.

The reason is not that one model remembered more. Sonnet 3.7 recorded things the game screen already showed every turn, such as the player's money and current location, so its archive was permanently going out of date and had to be constantly re-checked. Fable 5 used memory only for what the game never volunteered: strategy, plans and tested rules. Its top-level file held the current goal, a prewritten plan for the game's challenges, and its own rules for using memory.

The cost of the hoarding approach shows in the numbers. Around step 10,000, Anthropic observed Sonnet 3.7 start a fourth folder of "master" memory files. Over the following 17 steps it created six files, deleted nine others, and took no actions in the game at all. It was tidying its notebook instead of playing. The research also found that imposing fixed rules about what to save, such as "record only failures", stripped much of the memory benefit away from stronger models: freedom to judge salience beat prescriptive instructions.

Dreaming: the scheduled pass that repairs bad memory

Writing notes in the moment is a prediction about the future, and some predictions are wrong. The research documents one in detail: while navigating a cave system, an agent wrote down that a landmark building was on the east side of the mountain. It was on the west. Working from its own wrong note, the agent announced the dungeon behind it was complete, walked off a one-way ledge, and landed back where it had started. Without correction it made the same mistake three times, and the run ended without reaching its destination.

The fix Anthropic tested is not a better prompt at write time. It is a later pass. Running dreaming over the memory folder, the agent checked its note against the original session transcript, caught the error, and rewrote it; with the corrected note it navigated straight through [1]. The transcript is the ground truth, and the consolidation is when it gets applied.

This is not just research-stage tooling. Anthropic's Claude Cookbook publishes a working recipe in which a background workflow reads finished session transcripts and distils them into a wiki-style memory store that future sessions mount read-only, scheduled to run nightly from cron [2]. Its measured cost is modest: over about fifty long transcripts, roughly 12 MB of text, a dream cost about $8 at list price and took about ten minutes. The recipe's own guidance is worth quoting in spirit: memory is only as good as the transcripts, facts carry staleness notes, and anything that matters should be re-checked against its source.

When the memory representation beat search, and when it did not

The wiki experiments show where structured memory genuinely pays. Given an internal document corpus of about 350,000 files, roughly 420 million tokens, agents built a wiki over 128 sessions, and scores on a 25-question evaluation grew from 13% to 48% of graded criteria as the dreaming sessions scaled up [1].

The pattern in the gains matters more than any single figure. On questions that required finding information spread across several documents and applying judgement to it, scores reached 87% when the agent consulted its wiki against 73% with search alone. Combining the wiki with search lifted aggregate performance from 54% to 61% as the wiki grew, against 53% for search by itself, and saved about 20,000 tokens per question. But on tasks where every answer sat retrievable inside a single document, the wiki showed no clear gain. On one task it even grew larger than the documents it summarised, abstracting nothing: Anthropic compares it to a map so detailed it matched the empire it depicted.

Instructions shaped the outcome too. On a contract-review task, a wiki built without goal-focused instructions captured only 41 of 57 required facts, and a brief generated from it met 72% of grading criteria. After a reflection loop rewrote the instructions so each contract got its own page with terms quoted word for word, 100% of the criteria could be met [1]. The store did not just need more content. It needed to know, in advance, what the work would be judged on.

There is also a scaling discipline at the top of the hierarchy. Anthropic's own Claude Code loads only the first 200 lines, or 25 KB, of its memory file into any session, using it as an index rather than a record [1]. The injected layer must stay small, because it is paid for on every session and a wrong note there travels into all of them.

What this means for a business running agents

The operating lesson survives translation out of games and benchmarks. An agent's memory folder is a decision about what deserves to exist as a note, and the wrong default is to record everything, because state the environment already shows goes stale and judgement does not. If your agent has a memory directory today, the two questions from this research are: how many files does it hold, and how many of them say something the system of record would not show you anyway?

The second lesson is that memory needs maintenance with hindsight, not just capture in the moment. A scheduled pass that checks notes against the transcripts they came from is the difference between the corrected run and the one that walks into the same ledge three times. The mechanism is cheap; the discipline is the point. Anthropic's own framing of the open questions is candid: dreaming costs compute, no model can know in advance what future turns will need, and memory systems are still being worked out [1].

This maps closely onto what we already run. The agent that produces the drafts on this site keeps a deliberately small top-of-hierarchy file that works as an index, dates every daily note underneath it, and records in a state file which source items were processed and which expired unprocessed. Notes get corrected when a later pass finds they were wrong. That structure was not adopted from this research, but the research describes the same shape, and for the same reason: the memory that helps is the memory that carries judgement.

Sources

  1. [1] [1] Anthropic, "The components of agent memory", 6 November 2026, [2] Anthropic Claude Cookbook, "Dreaming with a dynamic workflow", 2 November 2026,