
Quick Answer
Agent memory gives Claude information it can retain across sessions, including knowledge gained from users, assigned tasks and prior interactions. Anthropic’s November 6, 2026 research describes how memory notes, selective guidance and dreaming can help Claude organize knowledge, correct inaccurate notes and build representations for future work, according to Anthropic’s Components of Agent Memory.
In This Article
- What This Means for Your Business
- What Does Agent Memory Include?
- How Does Claude Learn From Memory?
- Agent Memory Poisoning Proof of Concept and Claude Code Memory Research Agent
- How Does Dreaming Correct Agent Memory?
- What Does the Evidence Show?
- How to Keep Business-Agent Memory Accurate
- What to Do Now
- Key Considerations
- FAQ
- Related Coverage
What This Means for Your Business
For business owners, consider whether retained business facts remain current, whether public listings are consistent, whether agents recall relevant information, whether claims have supporting sources and whether the information is visible in AI answers. The findings point to a clear operating approach: retain durable facts separately from temporary conversation details, connect every retained fact to its source, record corrections and check public business information before it is published.
Anthropic’s Pokémon Red evaluation also shows why saving indiscriminately can hurt. In its one-run comparison, Claude Sonnet 3.7 created 413 files across 225 folders and earned one badge, while Claude Fable 5 wrote nine files in two folders, plus a top-level file, and earned four badges. These are research results from Anthropic’s game evaluation, not a direct test of business visibility.
What Does Agent Memory Include?
Agent memory is knowledge that Claude gains from experience, including user interactions and assigned tasks, and retains between sessions as context it can read later. Anthropic’s Team Engineering research group, in its November 6, 2026 paper, excludes changes to model weights, knowledge gained within a single conversation and compaction summaries that replace older turns.
Claude records work in logs and transcripts. It can tag salient events as they happen or write selected information to a persistent store such as a folder of memory notes. A retained note may hold a verified fact, a rule derived from an earlier task, a correction or information about a trap encountered during the work.
How Does Claude Learn From Memory?
On the Continual Learning Bench 1.0 database task, Claude received the correct answer after each attempt and could save a memory for future questions. If it answered incorrectly, it could record the failure, propose explanations, test them and convert the result into a future-answer rule.
Anthropic ran the task with Claude Managed Agents. Each question started in a fresh session, except that the agent could retrieve files from a memory directory that persisted between sessions. The paper’s model comparison shows how the systems differed:
| Model | Memory behavior reported by Anthropic |
|---|---|
| Claude Sonnet 4.6 | Generated theories about incorrect answers, but stored untested guesses. |
| Claude Opus 4.7 | Consolidated notes into one reference document and separated checked facts from unconfirmed information. |
| Claude Sonnet 5 | Marked answers as confirmed after grading and wrote corrections after errors. |
| Claude Opus 5 | Closed the learning loop with rules, supporting evidence, earlier answers and warnings about discovered traps. |
| Claude Opus 5.5 and Claude Sonnet 5.5 | Both closed the loop and captured validated rules. Their memory-assisted scores were about the same as Opus 5, while their baselines without memory were slightly higher. |
Merely telling Opus 4.7 and Opus 5 that a memory folder was available gave both a noticeable performance boost in Anthropic’s CL-Bench evaluation. Detailed fixed rules worked less well for these stronger models. A rule directing the agents to record only failures removed much of the performance benefit, while specific guidance helped older models more. Prompting Sonnet 4.6 to read its memory and test its theories nearly doubled its benchmark score. Sonnet 5 read its memory before answering nine out of 10 times without being prompted, compared with less than half the time for Sonnet 4.6.
Agent Memory Poisoning Proof of Concept and Claude Code Memory Research Agent
Consider reviewing persistent memory for signs of poisoning and separating trusted instructions from untrusted content.
Consider using a research agent to investigate topics, retrieve relevant context, store useful findings and improve through feedback and personal preferences.
How Does Dreaming Correct Agent Memory?
Dreaming lets Claude use idle time between interactions to review logs, reorganize context, correct mistakes, create notes and build new representations. Anthropic’s method checks a memory note against the original transcript before relying on it again.
One Pokémon Red run illustrates the mechanism. Claude incorrectly recorded the PokéCenter as being on the east side of Mt. Moon. It later marked Mt. Moon as completed and crossed a one-way ledge that returned it to its earlier route. Dreaming checked the location against the transcript, found the error and rewrote the note. With the corrected note, Claude returned to the cave entrance and eventually reached Cerulean City. With the uncorrected note, it returned to the earlier route three times and finished in Pewter City without reaching Cerulean City.
Dreaming can also create a wiki, a collection of pages written and maintained by Claude that links information about one subject to other notes. Anthropic explored this representation using roughly 350,000 internal files, or about 420 million tokens, from Stainless, a software company it acquired.
What Does the Evidence Show?
Anthropic’s Pokémon evaluation was a one-run comparison of Claude Sonnet 3.7 and Claude Fable 5 at roughly 10,000 game steps. Sonnet 3.7 wrote 413 files across 225 folders and earned one badge. Fable 5 wrote nine files in two folders, plus global.md, and earned four badges. Around step 10,000, Sonnet 3.7 started a fourth folder of master memory files; over the next 17 steps, it created six files, deleted nine and took no in-game action. Fable 5 focused on strategy, plans and tested rules, including guidance to vary timing after three or more failed Poké Ball attempts.
A separate mouse study by Wannan Yang and colleagues, published in Science on March 28, 2024, combined large-scale neural recordings with dimensionality-reduction techniques during maze traversals. The researchers observed sharp wave ripples after reward consumption and found that sleep ripples replayed sequences reactivated most often while the mice were awake. The study described awake ripples as a possible tagging mechanism for selecting experiences that undergo long-term consolidation. This neuroscience research supplied an analogy for Anthropic’s design, not an evaluation of Claude’s memory system.
The broader survey Rethinking Memory Mechanisms of Foundation Agents in the Second Half classifies foundation-agent memory by substrate, cognitive mechanism and subject. Its authors describe internal and external memory, with episodic, semantic, sensory, working and procedural functions, alongside agent-centered and user-centered approaches.
How to Keep Business-Agent Memory Accurate
- Separate durable business facts from information used only within one conversation.
- Attach each retained claim to the source record that supports it.
- Record corrections as new entries linked to the affected claim.
- Review stored entries for details that may have changed.
- Verify public business information across listings and before publication.
- Review retained instructions from untrusted content for signs of poisoning.
What to Do Now
- List the business facts your agents may reuse across sessions.
- Attach a source and last-reviewed date to each retained fact.
- Remove or separate conversation-only details from persistent memory.
- Check public business details across the listings your customers and AI systems may read.
- Review stored instructions and resolve conflicts against current source records.
Key Considerations
- Anthropic’s results come from CL-Bench and Pokémon Red, not a business-listing study.
- The Pokémon comparison used one run of each model at roughly 10,000 steps.
- Agent memory differs from weight updates, in-conversation learning and compaction.
- Consider evaluating persistent memory for poisoning risks separately from model performance.
- No official independent agent-memory leaderboard exists, according to the Agent Memory Leaderboard guide, and published scores depend on the answer model, judge, token budget and test harness.
FAQ
What Is Agent Memory?
Agent memory is knowledge Claude gains from users, tasks and prior experience that it retains between sessions as context it can read later. It excludes weight changes, knowledge gained only within one conversation and compaction summaries.
How Does Claude Decide What to Remember?
Claude tags salient events and can write selected information to persistent notes. In Anthropic’s tests, giving stronger models broad guidance and leaving the salience decision to them performed better than imposing fixed rules such as recording only failures.
Can Claude Correct an Incorrect Memory?
Yes. Anthropic’s “dreaming” process checks stored notes against original transcripts and can rewrite incorrect information. In the Pokémon Red example, a corrected location note helped Claude return to the Mt. Moon cave and reach Cerulean City.
What Risks Come With Persistent Agent Memory?
Review persistent memory for unexpected instructions and validate retained information against original transcripts.
How Should a Business Manage Agent Memory?
Retain durable facts separately from temporary conversation details, attach each fact to its source, record corrections, review stale entries and verify public business information before publication.
Related Coverage
- Anthropic Releases Claude Opus 4.7 With Extended Reasoning and Agent-Ready Architecture – BizScoreAI
- Claude Fable 5 Extended Through July 12: 10 Uses
- Deepseek V4.1-Flash Cuts AI Agent Memory Costs
Sources
See our editorial disclaimer for how our articles are produced.
Run a free scan to see your AI Visibility Score, SEO rating, and local citation accuracy.