The components of agent memory
by Anthropic
Anthropic's measured account of how agents write memory across sessions and how an offline 'dreaming' pass corrects it
Overview
The components of agent memory is an engineering essay published on Anthropic's engineering blog on November 6, 2026, written by Lance Martin. It defines agent memory as knowledge an agent gains from experience and keeps across sessions, which excludes weight updates, in-context learning and compaction. The essay then introduces "dreaming," an offline pass in which an agent rereads its own logs and transcripts to correct notes and build new representations, an idea it credits to Letta and UC Berkeley's sleep-time compute work. Sections cover memory as a prediction about what future sessions will need, how newer Claude models judge what is worth saving, how to prompt an agent to write memory, and four uses of dreaming: correcting wrong notes, building a linked wiki, promoting notes into always-loaded context, and focusing memory on a specific goal. The evidence is quantitative. On the CL-Bench 1.0 database task, a persistent memory directory raised Opus 5 from 0.564 to 0.828 mean reward, and prompting Sonnet 4.6 to read memory first moved it from 0.364 to 0.641. On a Stainless corpus of roughly 350,000 files, a Claude-written wiki plus search reached 61% of graded criteria against 53% for search alone and saved about 20,000 tokens per question. Claude Plays Pokemon runs compare a 413-file memory folder that stalled an agent with a nine-file folder that earned four badges. The experiments ran on Claude Managed Agents, and the essay ends with open questions and pointers to cookbooks on dreaming and knowledge wikis.
At a Glance
- Topic
- Agentic
- Level
- Intermediate
- Format
- Guide
- Cost
- Free
- Duration
- ~13 min read (about 3,300 words)
- Provider
- Anthropic
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓How Anthropic defines agent memory and why it excludes weight changes and compaction
- ✓Why writing a memory note is a bet on what future sessions will need
- ✓How to prompt an agent about the purpose of its memory folder, and why a fixed 'record only failures' rule cut gains
- ✓How an offline dreaming pass checks notes against transcripts to fix wrong memories
- ✓When a Claude-written linked wiki beats plain text search over a large corpus, and when it adds nothing
- ✓How to keep the always-loaded memory layer small by using MEMORY.md as an index
- ✓Why several agents writing to one shared table lost facts that one agent per contract kept
Highlights
- •Reports measured CL-Bench 1.0 results for six Claude models, with and without a memory directory
- •Includes failure cases, such as a wiki that grew to 1.77 times the corpus size without abstracting anything
- •Uses Claude Plays Pokemon runs to show how a cluttered 413-file memory folder slows an agent down
- •Ends with open questions on when to write memory and how to keep it current
Who It’s For
Best For
- ✓Engineers adding persistent memory to long-running or multi-session agents
- ✓Teams using Claude Managed Agents or Claude Code auto memory
- ✓Builders choosing between transcript search and a curated knowledge wiki for agent retrieval
Prerequisites
- •Experience building or running an LLM agent with tool use
- •Basic familiarity with context windows, compaction and retrieval
FAQ
What is The components of agent memory?
An Anthropic engineering essay on how AI agents should store what they learn across sessions. It is for engineers building long-running or multi-session agents. After reading it, you can design a memory folder, prompt an agent to use that folder well, and add an offline pass that corrects stale or wrong notes and builds a linked wiki.
Is The components of agent memory free?
The components of agent memory is free to access.
What level is The components of agent memory for?
The components of agent memory is aimed at a intermediate audience. Recommended background: Experience building or running an LLM agent with tool use, Basic familiarity with context windows, compaction and retrieval.
How long does The components of agent memory take?
Expect roughly ~13 min read (about 3,300 words). Most learners work through it at their own pace.
What will I learn from The components of agent memory?
You'll learn: How Anthropic defines agent memory and why it excludes weight changes and compaction; Why writing a memory note is a bet on what future sessions will need; How to prompt an agent about the purpose of its memory folder, and why a fixed 'record only failures' rule cut gains; How an offline dreaming pass checks notes against transcripts to fix wrong memories; When a Claude-written linked wiki beats plain text search over a large corpus, and when it adds nothing; How to keep the always-loaded memory layer small by using MEMORY.md as an index; Why several agents writing to one shared table lost facts that one agent per contract kept.
Topics
Sources
This page was written from 3 sources, 2 on domains other than anthropic.com.