Claude agent memory just got a clearer playbook. In a November 6 research post, Anthropic showed that agents remember best when the model decides what matters, then tidies its notes offline in a process it calls “dreaming.” In Anthropic’s tests, that lifted scores on a database task by roughly 0.2 to 0.3 points for newer models.
Here’s what you need to know before you build on it.
Quick Facts
These are the points worth screenshotting if you only have a minute.
- Published by Anthropic’s engineering team on November 6, 2026.
- Memory is split into four layers: transcripts, memory notes, memory representations, and injected context.
- “Dreaming” is an offline pass where Claude reviews past sessions and fixes its own notes.
- Sonnet 5.5 went from 0.644 to 0.833 on the CL-Bench database task with memory on.
The takeaway: memory isn’t just a bigger notebook. How the agent curates it matters more.
The four layers of Claude agent memory
Anthropic describes memory as a stack that sits between sessions. Each layer is more refined than the one below it.
- Transcripts: a searchable log of what happened.
- Memory notes: short observations Claude writes for itself.
- Memory representations: organized structures such as a wiki.
- Injected context: the small slice loaded at the start of every session.
The point of the stack is that an agent shouldn’t reread everything every time. It should load a little and know where to look for more.
What “dreaming” actually does
Dreaming is Anthropic’s name for a cleanup job that runs when the agent is idle. Claude reads old logs, corrects mistakes, rewrites notes, and decides what to preload next time. Anthropic links the idea to hippocampal replay in sleep and to earlier “sleep-time compute” research.
One example from the post shows why it matters. In a Pokémon Red test, a wrong note about a location sat in memory. After a dreaming pass fixed it, Claude reached Cerulean City at action 1,935. With the bad note left alone, it never got past Mt. Moon.
The numbers: how much memory helps
Anthropic ran the CL-Bench database task with and without memory. Newer models gained the most.
| Model | Without memory | With memory |
|---|---|---|
| Opus 5.5 | 0.655 | 0.838 |
| Sonnet 5.5 | 0.644 | 0.833 |
| Opus 5 | 0.564 | 0.828 |
| Sonnet 4.6 | 0.330 | 0.364 |
Sonnet 4.6 barely moved. Anthropic says older models needed extra prompting to read and test their own memory, and with that help it reached 0.641.
Why “only record failures” backfired
The most useful finding is a warning against over-controlling your agent. When Anthropic told models to record only failures, their gains shrank. Opus 5 lost 0.234 points and Opus 4.7 lost 0.150.
Stronger models did better when they judged for themselves what was worth keeping. Simple guidance on what memory is for beat step-by-step rules. In Pokémon Red, Claude Fable 5 earned four badges with nine files, while Sonnet 3.7 earned one badge with 413.
If you already juggle several agents, our guide to Mistral Managed Agents shows how another vendor handles agent setup. And if you want memory shared across chat apps today, see Notion shared memory for ChatGPT and Claude.
Real-world tests: wikis beat plain search
Anthropic also tried memory as a wiki that Claude builds for itself. On a roughly 420M-token body of material from Stainless, wiki-only answers rose from 13% to 48% over 128 dreaming sessions. Wiki plus search reached 61%, against 53% for search alone.
The wiki also saved about 20,000 tokens per question, which means lower cost as well as better answers. That matters if you pay per token to run agents all day.
A second test used Harvey’s merger questions. Multi-document questions scored 87% with a wiki, versus 73% with search alone. Per-contract pages lifted graded facts from 41 of 57 to 57 of 57.
Where you can try it
Anthropic ran its tests on Claude Managed Agents, which support a persistent memory directory. The post also links the Claude Platform memory tool, plus cookbooks for dreaming and knowledge wikis. In Claude Code, auto memory loads only the first 200 lines or 25 KB of MEMORY.md at session start.
That limit is a good reason to keep your top-level notes short and move detail into linked files.
Frequently Asked Questions
What is dreaming in Claude?
It’s an offline process where Claude reviews old logs, corrects mistakes, and rewrites notes for future sessions. It happens while the agent is idle.
Does more memory make an agent better?
Not by itself. Anthropic found that useful, well-curated notes matter more than saving every step.
Does Claude agent memory work on older models?
Less well. Sonnet 4.6 gained only slightly until Anthropic added prompting that told it to read and test memory.
How do I use memory in my own agents?
Start with the Claude Platform memory tool or Managed Agents, then add a dreaming pass once you have real session logs.
Our take: give the model a goal, not a rulebook
If you’re building agents, resist the urge to write long rules about what to remember. Tell the model what memory is for, keep the always-loaded notes small, and schedule a cleanup pass. The data suggests that’s where most of the gain comes from, though these are Anthropic’s own benchmarks, so test on your own tasks before trusting the numbers.
Leave a Reply