Does more agent memory make your AI agent worse?

Last Updated: 7 November 2026

What is agent memory?

Anthropic defines agent memory as knowledge Claude gains from experience, including its interactions with users and the tasks it is assigned, and retains between sessions as context it can read later. The definition deliberately excludes three other ways a model can change its behaviour: weight updates from training, in-context learning inside a single conversation, and compaction, where earlier turns are replaced by a summary so work can continue in a new context window. That boundary matters commercially, because only the first category is memory you can open, correct, delete and show to an auditor.

The working version, in most production systems, is unglamorous: a folder of plain text files with an index on top. Anthropic's own Claude Code auto memory keeps a MEMORY.md index plus one topic file per memory, and loads only the first 200 lines or 25 KB of that index into each session. Everything else is read on demand.

Why agent memory became a harness question

Through 2026 the harness, the scaffolding around the model, has mattered as much as the model itself. Anthropic published The components of agent memory on 6 November 2026, written by Lance Martin with figures by Monika Tuchowska, as an evaluation write-up rather than a product launch. It follows the May 2026 research preview of dreaming for Claude Managed Agents, the feature that lets an agent reorganise its own memory between sessions.

The reception is worth noting because it is thin. The Hacker News submission drew 6 points and no comments. The tweet that surfaced the post drew 16,956 views, 152 likes and 34 replies in its first day, from an account with 328,126 followers. There is no public rebuttal thread to weigh against Anthropic's numbers, so the honest read is that this is vendor-published evidence with unusually detailed failure cases, not an independently replicated result.

The five parts of a working memory system

Anthropic describes five components that do different jobs. Memory notes are written during a session, when the agent judges something worth keeping. Predictions are the forward-looking part: writing down a verified fact is a bet that a future session will rely on it without checking again. Dreaming is the offline consolidation pass, borrowed from sleep-time compute research at Letta and UC Berkeley. Memory representations, such as wikis, are maps of what the agent knows. Injected context is the small layer promoted to the top of the hierarchy and loaded into every session.

Each component fails differently. Notes go stale. Predictions can be wrong from the start. Dreams consume compute. Wikis can grow larger than the corpus they describe. Injected notes are read as trusted in every future session, so a single bad line has the widest blast radius in the system.

What the evidence actually shows

Anthropic ran five evaluations across three very different domains: a database question-answering benchmark, Pokemon Red, a 350,000-document internal corpus from Stainless, and two legal tasks from the Harvey LAB benchmark. The pattern is consistent. Memory helps most when the task requires rules, plans or cross-document judgement, and helps least when the answer is already visible in the environment or a single document.

Evaluation	Measured result	What it tested
CL-Bench 1.0 database task	Opus 5 reached 0.828 with memory (16 runs); Opus 4.7 reached 0.700 (3 runs); Sonnet 4.6 started at 0.330 and reached 0.641 once told to read memory and test its theories	Whether a model can build its own learning loop
Pokemon Red, about 10,000 steps in	Sonnet 3.7: 413 files in 225 folders, 1 badge. Fable 5: 9 files in 2 folders plus one index file, 4 badges	Judgement about what is worth saving
Stainless corpus, 25 questions	13% of graded criteria at 2 dreaming sessions, 48% at 128. Wiki plus search rose from 54% to 61%, against 53% for search alone	Whether a wiki becomes a usable map
Harvey LAB merger questions	87% with the wiki against 73% with search alone on the cross-document subset	Retrieval versus representation
Contract facts task	41 of 57 graded facts survived (72%) with one shared page. 57 of 57 survived (100%) after one owner per contract page and word-for-word quotes	Whether dreaming instructions target what matters

Where more memory makes an agent worse

Memory can make an agent worse in five specific ways: storing state the environment already shows, imposing fixed rules that override the model's judgement about salience, letting a wiki grow past the source material, promoting a wrong note into every session, and paying for consolidation runs that never earn their cost. Anthropic's evaluations show each failure with concrete numbers, and several are the kind a business team will reproduce in its first month of running a memory folder.

State is not memory. Sonnet 3.7 archived almost everything it saw: encounters, items, every non-player character, with a separate file for each battle. By step 10,000 it had 413 files across 225 folders and one badge. Over the next 17 steps it created six files, deleted nine and took no actions in the game at all. The game state changed constantly, so the archive was perpetually out of date, and the agent spent its budget combing through notes instead of playing.

Fixed rules backfire. When Anthropic told the models to "record only failures", Opus 5 and Opus 4.7 lost much of the boost memory had given them. The chart shows Opus 5 at negative 0.020 and Opus 4.7 at negative 0.234, meaning memory had become a drag. Stronger models did better with freedom to judge salience.

A wiki can abstract nothing. On one legal task the wiki grew to 68 pages, 255 links and 203 KB against a corpus of 115 KB, about 1.77 times the source material. Anthropic compares it to Borges's mapmakers, who drew a map so detailed it was the size of the empire.

Injected notes travel. A wrong note promoted to the injected layer is read as trusted in every session. Anthropic also repeats its Claude Code warning that longer instruction files reduce adherence, which is why the injected layer has to stay small.

Dreaming is compute. Correcting drift requires running consolidation jobs, and Anthropic is explicit that continuously checking every memory may not be practical. The sensible approach is to dream on the information most likely to change, at a cadence tied to how often it is read.

How dreaming works in practice

A dream is an asynchronous job that reads an existing memory store and between 1 and 100 past session transcripts, then writes a new store with duplicates merged, stale entries replaced and new insights surfaced. The input store is never modified, so you can review the output and discard it. Jobs typically run from minutes to a few hours depending on transcript volume, and are billed at standard token rates for the model you choose. Cost scales roughly linearly with the number and length of the sessions you feed in.

Two details matter for anyone planning a rollout. First, the instructions field steers synthesis rather than editing: asking it to focus on coding preferences works, asking it to change a specific sentence generally does not. Second, the output is an ordinary memory store, which means you can attach it, diff it against the input, or throw it away without touching production memory.

The memory hierarchy, and what to promote

Anthropic's hierarchy has three practical layers. The top layer is injected into every session and must stay small: Claude Code caps it at the first 200 lines or 25 KB of the index file. The middle layer holds topic files that are read on demand. The bottom layer is the transcript archive, which dreaming mines and which search can reach. The design rule is to keep detail out of the top layer and leave links behind, so the agent knows what exists without paying for it every turn.

Promotion is where the leverage sits. Anthropic's Pokemon example shows why: the useful notes were strategy, plans and tested rules, things that exist only if the agent writes them down. Facts the screen already shows every turn, such as the player's money or location, are redundant in memory and go stale fastest.

Six rules to copy from this research

The post is a research write-up rather than a playbook, but the evidence supports six operating rules any team building agents can apply this quarter: cap the injected layer, write judgement instead of state, say what memory is for, give each page one owner, dream on a cadence tied to volatility, and review every output store before attaching it. Each rule maps to a measured failure in Anthropic's evaluations.

1. Keep the injected layer under about 200 lines. Treat it as an index of what the agent knows and where to find it, not as a knowledge base. Anthropic's own tooling enforces the same limit.
2. Write rules and plans, not state. Store strategies, tested rules and warnings about traps. Do not store what the environment already shows, because it changes and the note will be wrong within hours.
3. Say what memory is for, not how to use it. A sentence telling the agent a memory folder exists was enough for the stronger models. Prescriptive rules such as "record only failures" measurably hurt.
4. Give every page one owner. Parallel writers on a shared page overwrote each other and lost 16 of 57 contract facts. One writer per page, quoting source language word for word, took the task to 100%.
5. Dream on a cadence tied to volatility. Consolidate the notes most likely to drift, on a schedule set by how often they are read. Do not try to re-check everything continuously.
6. Review the output store before you attach it. The API leaves the input untouched for exactly this reason. Diff the two stores, keep what improved, discard the rest.

What it costs, and what it does not fix

Dreaming is in research preview on Claude Managed Agents and requires two beta headers. It is not generally available, and the supported-model list excludes several current models, including Haiku 5.5 and Opus 4.6. Memory stores carry hard limits: 100 kB per memory, about 25,000 tokens, 10,000 memories per store, and a maximum of 8 stores per session.

The compliance position deserves a plain sentence. Managed Agents, including memory stores and dreams, are not zero data retention eligible and not HIPAA eligible, and session transcripts persist until you delete them. Memory stores also attach with read and write access by default, which means untrusted input can write content that later sessions read as trusted memory. For client records, health data or anything covered by Australian privacy obligations, use read-only stores for reference material and keep write access narrow.

None of this replaces evaluation. Anthropic's numbers are its own, on its own harnesses, with small run counts on several comparisons. The transferable finding is structural: memory pays when it carries judgement, and costs when it carries state.

Frequently asked questions

What is agent memory?

Agent memory is knowledge an AI agent gains from experience and keeps between sessions as context it can read later. It is distinct from model weights, in-context learning within one conversation, and compaction, which summarises earlier turns. Memory is the only one of these layers a person can open, edit, delete and audit.

What is dreaming in Claude?

Dreaming is a consolidation pass where Claude reviews past session transcripts alongside an existing memory store and writes a new, reorganised store. It can correct mistakes, merge duplicates, build wiki-style representations and promote notes into the injected layer. It is a research preview for Claude Managed Agents, runs asynchronously, and is billed at standard token rates.

Does more agent memory make an AI agent better?

Not automatically. In Anthropic's Pokemon evaluation, a model that wrote 413 memory files earned one badge while a model that wrote nine files earned four. Memory helps when it captures rules, plans and cross-document judgement, and hurts when it captures state the environment already shows or when fixed rules replace the model's own judgement about salience.

How much does it cost to maintain agent memory?

Capture is cheap; maintenance is not. Dreaming is billed at the standard token rate for the model you select, and cost scales roughly linearly with the number and length of session transcripts you feed in. Anthropic recommends starting with a small batch of sessions and scaling once curation quality is proven.

Is Claude's agent memory safe for regulated data?

Not by default. Claude Managed Agents, including memory stores and dreams, are not zero data retention eligible and not HIPAA eligible, and transcripts persist until deleted. Memory stores attach read-write by default, so prompt injection can write content that later sessions treat as trusted memory. Use read-only stores for reference material and keep write access narrow.

What should a business team do first?

Start with a small, human-readable memory folder: one index file under 200 lines, one topic file per subject, and a rule that only strategies and tested rules get written down. Then add a consolidation pass on the notes that change most often, and review its output before attaching it to any production session.