- Anthropic defines agent memory as knowledge Claude gains from experience and keeps between sessions as context it can read later. Weight changes, in-context learning and compaction are excluded.
- In its tests, newer Claude models wrote better notes, and on Pokémon Red a model with 9 memory files beat one with 413 files.
- Dreaming means the agent reviews its logs between sessions to fix wrong notes and build new memory representations such as wikis. In one Pokémon test it fixed a bad note and the run got past the point where uncorrected runs failed.
What happened
Anthropic published a research post on the components of agent memory. It defines agent memory as knowledge Claude gains from experience, including user interactions and assigned tasks, that it keeps between sessions as saved context. It leaves out weight changes, in-context learning, and compaction.
The post covers two main pieces. The first is memory notes, which Claude writes in-session based on what it judges to be salient. The second is dreaming, where Claude uses idle time between interactions to reflect on logs, correct mistakes, write new notes, and build new representations such as wikis. The idea builds on sleep-time compute from researchers at Letta and UC Berkeley.
Anthropic backed this with tests. On Continual Learning Bench (CL-Bench) 1.0, a database question task run with Claude Managed Agents, Opus 5 was the first model to close the learning loop, meaning its notes held rules plus how it knew they were true. On Pokémon Red, Sonnet 3.7 wrote 413 memory files across 225 folders and earned one badge. Fable 5 wrote nine files in two folders plus one global.md and earned four badges. The wiki test on about 350,000 Stainless files (roughly 420M tokens) is cut off in the text I have, so those results are not confirmed here.
What this means for founders
- Less prompt engineering for strong models. Anthropic says it is enough to tell Claude a memory folder exists. For Opus 4.7 and Opus 5 that gave a noticeable boost on CL-Bench.
- Strict rules can backfire. Telling Opus 5 and Opus 4.7 to "record only failures" cost them much of their memory gain. Be careful before you hard-code what your agent may save.
- More memory is not better memory. Sonnet 3.7's huge archive went stale and it spent steps reorganizing instead of playing. Store what the agent has to work out itself (strategy, plans, tested rules), not what it can read off the screen every turn.
- Older models need more steering. Prompting Sonnet 4.6 to read its memory first and test its theories nearly doubled its benchmark score. If you run an older or cheaper model, add that guidance.
- Bad notes get trusted later. A wrong note can be read and relied on in a future session. Keep transcripts so a review pass can check notes against ground truth. Costs for dreaming are not stated in the source.
What the Launch Llama directory shows
Live numbers from Launch Llama for listings that describe themselves as "agent memory", "ai agent framework", "long-term memory" tools, as of 2026-11-08.
| Signal | Launch Llama data |
|---|---|
| Listings in this space | 3 |
| New in the last 30 days | 0 |
| With a free or freemium plan | 100% |
| Most upvoted by founders | Kimi AI - Now with K2.5 (336 upvotes) |
Launch Llama has only 3 directory listings matching this space and none went live in the last 30 days, though all of them have a free or freemium plan, so dedicated agent memory tools look thin and cheap to try.
How to try it today
- Give your agent a persistent memory directory that it can read and write between sessions. Anthropic used a mounted directory with Claude Managed Agents.
- Start with simple guidance. Tell the agent the folder exists and when to save notes (when a person asks, or when something is clearly worth remembering). Do not prescribe what to record.
- If you use an older or weaker model, add explicit steps to read memory before answering and to test its own theories about mistakes.
- Save session transcripts or logs, and run a separate review pass between sessions where the agent checks its notes against them and rewrites wrong ones.
- Look at the memory folder after a few runs. If it is full of per-event files or facts the agent could read directly, tighten the guidance toward strategy, plans and tested rules.
A starting prompt based on the post's advice (my wording, not Anthropic's):
A memory folder is available at /memory. Read it before you start each task. Save a note when a user asks you to, or when something is clearly worth remembering for future sessions. Prefer rules you have verified, and say how you verified them.
Related tools on Launch Llama
Tools from the Launch Llama directory that fit with this news:
| Tool | What it does | Why it matters here |
|---|---|---|
| Hosted Agents in Cluing | For teams: Cluing hosts collaborative AI agents that learn from your saved knowledge and execute tasks together, so your team builds institutional memory without technical setup. | Cluing hosts agents that learn from saved knowledge so teams build institutional memory, which is a hosted take on the cross-session memory idea in the post. |
| Mastra Factory | For AI developers: Mastra Factory builds AI agents with workflows, memory, and evals so you ship production apps from issue to deployment faster. | Mastra Factory lists memory and evals among its agent-building features, the same two areas Anthropic tests in this post. |
| Aramb | For founders: Aramb is an OS for AI agents where you build, launch, and monetize custom agents in 20 minutes with one API covering runtime, memory, browser, tools, and billing. | Aramb offers one API covering runtime, memory, browser, tools and billing for custom agents, so memory comes bundled rather than built from scratch. |
| AgentsKit | For developers: AgentsKit provides open-source TypeScript components for building AI agents so you avoid vendor lock-in and swap layers freely. | AgentsKit provides open-source TypeScript components for building agents, which lets you swap in your own memory layer. |
Browse all tools or list your own tool.
Llama's take
I would try this now if you run agents across many sessions and your memory setup is a pile of hand-written rules. The cheapest test is to loosen your prompt and see if a strong model picks better notes than your rulebook does. The post suggests it might.
I would wait on dreaming if your agents are short-lived or stateless. It adds a second pass and a transcript store, and the source gives no cost or latency numbers. Also keep in mind that these are Anthropic's own evaluations on its own models, and I only have a truncated version of the post.
The one thing to watch is the wiki result on the Stainless documents. That is where memory structure at scale gets tested, and the text I have cuts off before the findings.
FAQ
What are the components of agent memory?
In Anthropic's framing, agent memory is knowledge gained from experience and saved between sessions as context the agent can read later. The post focuses on memory notes written in-session, judgment about what is salient, and dreaming, where the agent reviews its logs to correct notes and build new representations like wikis.
What is dreaming in AI agent memory?
Dreaming is a process where an agent uses idle time between interactions to reflect on its logs or transcripts. It can fix mistakes, create notes, build new memory representations, or shape what gets preloaded into future sessions. In one Pokémon test, it caught a wrong location note by checking the transcript.
Should I tell my agent exactly what to save to memory?
For stronger Claude models, Anthropic found freedom to judge salience worked better than fixed rules. A 'record only failures' rule cost Opus 5 and Opus 4.7 much of their memory gain. Older models like Sonnet 4.6 benefited from explicit steering, such as reading memory first and testing theories.
Sources
- The Components of Agent Memory (www.anthropic.com)
- Hacker News discussion
