Anthropic's Agent Memory Study Puts Opus 5.5 at 0.838 With Memory and 0.655 Without

Anthropic published a research post on Nov. 6 that measures how much a memory folder, a wiki and an offline "dreaming" pass lift Claude agents, with gains from 0.034 to 0.317 points on one benchmark task. The page credits Lance Martin as writer and lists no job titles.

The post, The components of agent memory, defines memory as knowledge gained from experience and kept between sessions. It excludes weight changes, in-context learning and compaction. It names five parts: transcripts, memory notes, memory representations such as wikis, injected context, and dreaming, the offline process that corrects and consolidates the other four.

CL-Bench: six models on one database task

Anthropic reports mean reward on the database task of CL-Bench 1.0, with and without memory. Anthropic ran the tests.

Model	Without memory	With memory
Claude Opus 5.5	0.655	0.838
Claude Sonnet 5.5	0.644	0.833
Claude Opus 5	0.564	0.828
Claude Opus 4.7	0.383	0.700
Claude Sonnet 5	0.306	0.610
Claude Sonnet 4.6	0.330	0.364

Sonnet 4.6 reached 0.641 only after the prompting was refined. The post said Sonnet 5 read its memory before answering 9 of 10 times, against under half for Sonnet 4.6. Telling a model to record only failures cut the memory gain by 0.234 on Opus 5, 0.150 on Opus 4.7 and 0.020 on Sonnet 4.6.

Pokémon Red and 413 memory files

Anthropic used one run per model on Pokémon Red and calls the results illustrative. Sonnet 3.7 wrote 413 memory files across 225 folders and earned one badge. Fable 5 wrote nine files in two folders plus a global.md of about 43,000 characters, against about 978,000 for Sonnet 3.7, and earned four badges.

The post said "Sonnet 3.7 assumed everything might matter and wrote to memory indiscriminately." Around step 10,000 it spent turns reorganising memory with no game actions.

Wiki versus search on 350,000 files

On a Stainless corpus of about 350,000 files and roughly 420 million tokens, Anthropic ran 25 questions after 128 dreaming sessions. A wiki alone rose from 13% to 48% of graded criteria. Wiki plus search reached 61%, against 53% for search alone, and saved about 20,000 tokens per question.

1. Wiki plus search61 %
2. Search alone53 %
3. Wiki alone48 %

Source: Anthropic, The components of agent memory, Nov. 6, 2026

The gains did not hold everywhere. On 24 Harvey LAB merger questions the wiki scored 87% against 73% for search on multi-document items. On one brief task it showed no clear gain, and on another the wiki grew to 1.77 times the size of the 115 KB corpus. Contract facts survived at 41 of 57 (72%) with a shared table and 57 of 57 with one page per contract and verbatim quotes.

Letta's earlier sleep-time compute work

The post lists sleep-time compute from Letta and UC Berkeley as related research. Letta's write-up, dated April 21, 2025, credits Kevin Lin, Charlie Snell, Yu Wang, Charles Packer, Sarah Wooders, Ion Stoica and Joseph E. Gonzalez. It says "sleep-time compute creates a Pareto improvement in model performance," tested on the AIME and GSM math benchmarks.

Those are different tasks from Anthropic's, so the numbers cannot be compared. The authors of the Anthropic post list costs: dreaming takes more compute, an injected note that is wrong carries into every session, and Claude Code loads only the first 200 lines or 25 KB of MEMORY.md.

Anthropic's other research is covered in our reports on its 1.35x to 2.02x research speedup study and on the METR review of its research threshold.

The post says the cadence and scope of dreaming are open questions.