AI NewsProductivityAnnouncement
Anthropic published agent memory results from Claude Opus 5.5 and Fable 5
Anthropic published a four-layer memory design for Claude agents and benchmarks on CL-Bench, legal contracts and Pokémon Red.
Image: Anthropic
Why it mattersA memory layer is now something a buyer can ask a vendor to measure. Anthropic published the numbers and the cookbook, so teams can run the same test on their own data.
An AI agent that forgets everything at the end of a chat cannot learn from its own mistakes. Anthropic published a research post about this. The post puts numbers on what good memory is worth to a Claude agent. It also sets out a four-layer design. Lance Martin wrote it. The code and example notebooks are on Anthropic's documentation site. The agents run on Claude Managed Agents.
Anthropic says an agent's memory has four layers. Transcripts are the searchable session logs. Memory notes are short observations Claude writes during a session. Memory representations are wikis that link notes together on a topic. Injected context is what gets pre-loaded into the next session. Between sessions, Claude runs an idle process Anthropic calls "dreaming". It reads the transcripts, fixes mistakes, and moves what matters up the layers.
The CL-Bench numbers
Anthropic ran Continual Learning Bench 1.0 on the same models with and without a memory store. Each question started in a fresh session and could read files from a memory directory that survived between sessions. By Anthropic's figures, Claude Opus 5.5 went from a reward of 0.655 without memory to 0.838 with it. Sonnet 5.5 went from 0.644 to 0.833. Opus 5 went from 0.564 to 0.828. Sonnet 4.6 went from 0.330 to 0.364.
Stronger models got more from memory. Opus 4.7 improved by 0.317 points. Sonnet 4.6 improved by 0.034. Anthropic says this is because a stronger model better judges what is worth saving in the first place.
Pokémon and a wrong note
Anthropic also compared two models about 10,000 steps into 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 a single top-level global.md, and earned four badges. Anthropic reads this as Fable 5 being more selective about what it saves.
Dreaming is the part that catches a mistake the first write got wrong. In Pokémon Red, Claude fell down a one-way ledge in Mt. Moon and ended back where it started. It then wrote "Mt Moon DONE" in its notes. After a dreaming session read the transcript and the next session read the corrected notes, Claude avoided the ledge. Memory alone did not fix the error.
The wiki test
For a wiki benchmark, Anthropic used a company-acquisition task from its LAB benchmark. It gave Opus 5 a set of documents and let it build a wiki over many dreaming sessions. Claude then answered 25 questions against the wiki. The number of dreaming sessions grew from two to 128. The share of graded criteria the wiki got right went from 13 percent to 48 percent.
A team building a Claude agent today can read the cookbook and copy the structure. The more useful question for a buyer is which number a vendor will report. Anthropic has set a template. Name the task. Run with and without memory. Say which model wrote which files.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
