Agent memory explained: Anthropic's components of agent memory, dreaming, and what the tests showed

A language model forgets everything when a session ends. Each new session starts from the system prompt and whatever you put in its context window, which is the text the model can see at one time. If an agent learned yesterday that a database column is named differently from what the docs say, it will make the same mistake today unless that lesson was written down somewhere it can read.

Anthropic's post defines agent memory as exactly that written-down part: knowledge Claude gains from experience, from users and from tasks, kept between sessions as context it can read later. In practice this is usually a folder of plain files that the agent can list, read and edit.

The post is careful about what memory is not. It is not a change to the model's weights, the numbers set during training. It is not in-context learning, which only lasts for one conversation. And it is not compaction, where Claude summarises its own older turns to keep working in a fresh context window.

Read next: Agent memory, defined in one page

While it works, the agent can tag important events in its session log or write them into a memory folder as notes. The hard part is judging, in the moment, what a future session will need. Anthropic calls this salience: picking out what matters before you know what comes next.

The post shows how much that judgement has improved across Claude models. On a database question task from CL-Bench 1.0, a benchmark for agents that learn across long tasks, each question started in a fresh session that could only read a memory folder. Claude Sonnet 4.6 left a scatter of files full of untested guesses. Claude Opus 5 was the first to close the loop: its notes held a rule, how it knew the rule was true, earlier answers and warnings about traps. Opus 5.5 and Sonnet 5.5 did the same.

Pokemon Red showed the same thing from another angle. About 10,000 steps into the game, Claude Sonnet 3.7 had written 413 memory files across 225 folders, one for nearly every battle and character, and had earned 1 badge. Claude Fable 5 had written nine files in two folders plus one global.md file, and had earned 4 badges. Sonnet 3.7 stored things it could read off the screen every turn; Fable 5 stored only strategy, plans and tested rules, the things that disappear unless written down.

Read next: The context window: what a model can see at one time

Notes written in the moment can be wrong, or go stale as the world changes, and a later session will trust them anyway. Dreaming is Anthropic's fix. Between sessions, a separate job reads the memory folder next to the raw transcripts of past sessions, and rewrites the memory. The name comes from how the human brain replays recent experiences during sleep. The idea builds on sleep-time compute, proposed by researchers at Letta and UC Berkeley.

The post describes five jobs a dreaming pass can do: correct mistakes by checking notes against what really happened, build new structures such as a linked wiki, promote a few notes to the top layer that loads in every session, focus memory on one goal, and fill gaps that no single session could see.

The clearest example is from Pokemon. Claude wrote down that a Pokemon Center was on the wrong side of a mountain, walked off a one-way ledge and blamed something else. In the run with the wrong note, it fell back three times and never reached the next city. After a dreaming pass checked the note against the transcript and rewrote it, Claude went straight back into the cave and reached Cerulean City.

Some notes are useful in every session, so the agent harness loads them into the context at the start, before the agent asks. Anthropic calls this promoting notes up the memory hierarchy. It is powerful, because the agent does not even need to know a fact exists to benefit from it.

It is also expensive and risky. Every loaded note costs tokens in every session, a wrong note is carried into every session, and loading a note does not guarantee the model acts on it. The post points to Claude Code's own documentation, which warns that longer instruction files reduce adherence.

So the top layer works like an index. Claude Code's auto memory loads only the first 200 lines or 25 KB of its MEMORY.md file, which tells Claude what it knows and where to find the rest. A large part of dreaming's job is keeping that layer short by moving details down into linked files.

Read next: Context rot: how long context degrades an agent

Anthropic lists several problems it has not solved. Writing notes during a task pulls the agent's attention away from the task, while dreaming in the background means its output is not available right away. Dreaming costs compute, so checking every memory all the time is not practical; the post suggests focusing on information that is likely to change, or tuning how often dreaming runs.

A wiki is not always worth it. On one legal brief task, most answers sat in a single document and a text search found them, so the wiki showed no clear gain. On another, the wiki grew larger than the documents it summarised. The post compares it to the map in a Borges story that was as large as the empire it described.

In product terms, Dreams is a feature of Claude Managed Agents and is in research preview. According to the docs, a dream takes an existing memory store and 1 to 100 past sessions, and writes a new, reorganised store without changing the input, so you can review the result and discard it.

Read next: The AI Engineering course: agents, memory, context and evals in production