Michael JohnsonIf an AI agent remembers a false fact, it may confidently repeat the same mistake in its next session. In research published by Anthropic on November 6, 2026, the company describes a way to review records between tasks: the agent compares them with logs of previous work, corrects inaccuracies, and combines useful information. The company calls this background process “dreaming.”
The practical takeaway for people designing long-running agents is that memory should not be judged by the number of saved notes, but by whether the system can check their provenance, update information that has changed, and find what it needs in a large archive. Anthropic’s experiments show both the value of this approach and the conditions under which memory starts to get in the way.
What the researchers mean by agent memory
In its description of memory components, Anthropic defines memory as knowledge from past interactions and tasks that persists across sessions as context available to the agent. It is a separate layer of the system: notes, files, indexes, and other external representations—not changes to the model’s parameters or an ordinary brief summary of the current conversation.
The system has several layers. A log or transcript records events; notes capture individual observations; linked representations, such as a wiki, organize information by topic; and a small top layer can be loaded in advance into a new session. The agent has to decide what to keep, what to check later, and what to show the next run immediately.
That last decision is particularly difficult: an event’s importance often becomes clear only after the agent sees what happens next. Anthropic therefore explored separating in-session work from background processing. While working, the agent leaves logs and notes; later, a separate pass can review them, compare sources, and update the memory structure.
How the mechanism was tested
Rather than using one general test, the company described several tasks: questions about a knowledge base, playing Pokémon Red, creating a wiki from corporate documents, and preparing a legal brief on contract terms. This mix tests different memory functions: retaining rules across sessions, selecting useful details, correcting errors, and combining information from multiple documents.
In one knowledge-base task, each question began in a new session. Only the memory folder persisted between sessions, and the model could refer to it. The researchers compared performance with and without memory and also observed how models created records. Anthropic notes an important distinction: a note containing a verified rule and an explanation of its provenance can help with the next question, whereas a collection of unverified guesses does not create a reliable learning loop.
In a separate experiment, an agent built a wiki from about 350,000 internal documents, or roughly 420 million tokens. Then 128 consecutive sessions added to and checked pages, corrected links, and merged duplicates. The evaluation consisted of 25 questions about the corpus. This was a narrow experimental task with a defined document set, not a universal assessment of memory for all agents.
What the experiments showed
In the knowledge-base task, memory improved the average result for several models. For example, Opus 5.5’s average share of completed criteria rose from 0.655 without a store to 0.838 with one; Sonnet 5.5 rose from 0.644 to 0.833. Performance nevertheless depended on how the model used its records: Claude Sonnet 4.6 often did not read the memory files before answering, while Opus 5.5 and Sonnet 5.5 were better at formulating verified rules and using them later.
The Pokémon experiment showed the cost of excessive archiving. In one run, Sonnet 3.7 created 413 files in 225 folders, recording many encounters, items, and characters. Much of this information quickly became outdated: the game’s position and state changed, and the agent spent actions sorting and restructuring its records. In a separate comparison of memory at around the 10,000-step mark, Fable 5 used nine files in two folders and a separate general file; according to Anthropic, Fable 5 earned four badges, while Sonnet 3.7 earned one.
A particularly clear example involves an incorrect note about the location of a Pokémon Center near Mt. Moon Cave. The agent assumed it had already passed through the cave and kept returning to the previous route. A background pass checked the note against the action log, found the error, and rewrote it. The agent then returned to the entrance and reached Cerulean City; the version with the uncorrected record remained stuck in the same area. Here, “dreaming” worked as a check of the history, not as free-form generation of new facts.
In the large-wiki experiment, the score against the 25-question criteria rose from 13% to 48% as the wiki was built. In a separate comparison, the wiki without search scored 48%; using the wiki together with search changed the result from 54% to 61%, while search without the wiki scored 53%. In this experiment, the wiki served more as a map of connections and a thematic layer over search than as a replacement for the original documents.
Why more memory is not always better
A record is a prediction about the future: the agent assumes that what it saves will be useful again. That prediction may become outdated, turn out to be wrong, or duplicate information available directly in the environment’s current state. In a game, the character’s position can be seen on screen, while a derived strategy or verified rule cannot be obtained in the same way without reasoning again.
The document experiment revealed another risk: structural errors. Several agents were consolidating contract terms onto a shared page at the same time and overwrote parts of the table. As a result, the wiki contained 41 of 57 verifiable facts, and the first brief met 72% of the criteria. After each contract was given its own page and important terms were copied verbatim, the brief reached 100% of the criteria. For legally significant wording, preserving the quotation and a link to the original document proved more useful than an abstract summary.
Finally, building a wiki is not always worth the cost. In a task where answers could be found in a single document, Anthropic saw no clear improvement over ordinary text search. In another test, the resulting wiki was larger than the original files. These results point to a straightforward rule: a complex structure is useful when questions regularly require connecting information across sources; for simple searches, it can add maintenance without noticeable benefit.
What can be applied to agent design
The results support a practical architecture with three parts. First, preserve primary logs so notes retain a verifiable foundation. Then create compact records of conclusions, rules, and plans—with an indication of where they came from and whether they have been verified. Finally, periodically run a separate pass that looks for contradictions in those records, corrects links, and rebuilds the index without replacing the original material with its own retelling.
For information that changes, it is useful to store a date or context indicating when it is current and to check the record when it is reused. For durable knowledge, preserve the source and the supporting evidence. And only a brief index should be loaded into each session’s general context: Anthropic separately notes that loaded instructions use up the context window, and an incorrect note automatically added to every run can spread an error repeatedly.
The study’s main lesson is that agent memory becomes useful through selection and maintenance, not accumulation. An archive preserves history, search finds relevant passages, structured notes connect recurring rules, and background checks trace conclusions back to primary evidence. That verifiable link is what turns past experience into a resource for the next session.
