Anthropic: Selective notes beat full logs for Claude agent memory, with 'Dreaming' fixing errors
Anthropic published research on how Claude retains knowledge across sessions using note files. Its tests show selective notes outperform exhaustive archives, and a post-session process called Dreaming can correct faulty notes. All results come from Anthropic's own tests.
Anthropic says an agent's memory works better when it records less. In a research post, the company reports that Claude agents that write selective notes outperform those that log everything, and that a post-session process called "Dreaming" can correct faulty notes after the fact. The model's weights stay unchanged throughout: Claude writes notes to a folder and reads them back in later sessions. All results come from Anthropic's own tests, as reported by The Decoder.
The selection problem
According to Anthropic, the core challenge is selection. The agent must decide in the moment what might matter later, before it knows what comes next, so every note is a bet on the future.
Pokémon Red test: 413 files vs. a handful
In a Pokémon Red test, the older Claude Sonnet 3.7 treated everything as potentially important. Over roughly 10,000 game steps it created 413 files in 225 folders, often one per battle and item. Because game state keeps changing, the archive went stale quickly. At times the model only reorganized its notes instead of playing, and it finished with one badge.
The newer Claude Fable 5 limited its memory to a few files and earned four badges. It did not record information already visible on screen, and instead wrote down plans, strategies and rules it had tested. Anthropic's comparison covers only a single playthrough per model.
Instructions: rigid rules hurt stronger models
Anthropic recommends explaining the purpose of memory to Claude and letting the model decide how to use it. In one benchmark, rigid instructions such as "only record mistakes" wiped out most of the advantage for stronger models, according to Anthropic. Older models benefited from clear instructions, such as being told to read their notes before every response.
Dreaming: reviewing old transcripts
The larger risk, per Anthropic, is faulty notes the agent trusts blindly. Dreaming, which builds on research from Letta and UC Berkeley, addresses this. During idle periods, Claude goes through transcripts of earlier sessions, checks its notes against them, and builds the upper layers of its memory.
In the Pokémon test, Claude had noted the wrong side of a mountain for a landmark. It fell into the same trap three times, wanted to give up, and missed its goal. After a Dreaming pass, Claude caught the error against the transcript and found the right path. It reached Cerulean City past Mt. Moon only with the corrected note.
Wiki test: Stainless documents and legal questions
Anthropic says Dreaming can also organize larger knowledge bases. Claude built a linked wiki from about 350,000 internal documents from Stainless, a software company Anthropic acquired. Dreaming generates new pages while fixing gaps, duplicates and broken links. Results were best when Claude combined the wiki with search, so the wiki complemented search rather than replacing it.
In a separate legal test, the wiki helped most on questions whose answers were spread across multiple documents. Claude scored 87 percent with the wiki and 73 percent with search alone. On simple lookup tasks the wiki performed roughly on par with search, and in one case it grew larger than the source documents.
Open problems
Anthropic lists several limits:
- Taking notes mid-task distracts the agent from its work.
- Dreaming costs extra compute, and results arrive with a delay.
- Wikis and notes need regular upkeep, or errors and hallucinations accumulate.
- Automatically loading too many memories into every session consumes context and passes any bad note to every future session.
Background
Anthropic introduced Dreaming in May as a research preview for Claude Managed Agents, its cloud platform for AI agents. The service reads an existing memory store and up to 100 past sessions, removes duplicate and outdated entries, and creates a reorganized store while leaving the original untouched. At launch it supported only Claude Opus 4.7 and Sonnet 4.6.
The Decoder notes that others use the same term. OpenAI calls ChatGPT's memory system, overhauled in June, "Dreaming"; it maintains a user profile in the background and corrects outdated information. According to OpenAI, this raised temporal accuracy from 52.2 to 75.1 percent. Researchers at Google and Google DeepMind use dreaming to improve an agent's search strategy, working through alternatives from stored results without calling the model again. Google has also built WikiSkill, a wiki for agents that, like Anthropic's system, keeps raw transcripts, distilled knowledge and specific instructions in separate layers.
What this means
Anthropic's findings shift the memory question from capacity to curation. In its Pokémon test, the model that saved the most (413 files) did worse than the one that saved a few files of tested rules. The Dreaming results suggest that reviewing notes against original transcripts is a practical way to catch errors an agent would otherwise repeat. The evidence has limits: the Pokémon comparison is one playthrough per model, and every test was run by Anthropic. The costs Anthropic names, extra compute and delayed results, apply to any team adopting a Dreaming-style pass. The instruction findings also matter for builders: per Anthropic, stronger models do better when told why memory exists than when given rigid rules, while older models still benefit from explicit instructions.
Related Articles
Claude can now build HTML-in-Canvas sites, needs Chrome or Edge 157 or Claude desktop 5.11492
Claude can now build web pages that embed real HTML elements such as text and buttons inside canvas graphics and 3D effects. The feature works in Chrome and Edge from version 157 and in Claude's desktop apps from version 5.11492. Users must request HTML-in-Canvas explicitly in their prompt.
Anthropic: Single-Turn Safety Tests Show 94% to 98% Harmless Rates Across Recent Models
Anthropic published recommendations for more reliable AI safety evaluations on November 13, 2026. It reports aggregate harmless-response rates of 94% to 98% for every recent model on its single-turn API tests, and says multi-turn evaluations offer more variation.
Claude Duet agent, in research preview for US Pro, Max and Ultra users, is useful but jumpy, The Verge finds
Claude Duet, Anthropic's AI agent for professionals and small business owners, is a research preview limited to US Pro, Max and Ultra subscribers, with a waitlist. In The Verge's hands-on, it connected information across email, Slack and Google Docs, but misjudged urgency and timing.
Anthropic launches Claude Meetings beta, putting Claude in video calls for Team and Enterprise plans
Anthropic began rolling out Claude Meetings in beta on November 12, 2026 for Team and Enterprise plans. Claude joins video calls as a visible participant that answers questions and turns discussions into notes, action items and editable documents.
Comments
Loading...