DEV Community

News
News

Posted on

Anthropic says agent memory works best when the agent knows what to forget

Infographic cover

๐Ÿ“… November 16, 2026 โ€ข The Decoder

Anthropic has published a research post called The components of agent memory by Lance Martin, dated November 6, 2026, and The Decoder covered it on November 16. Agent memory, defined as knowledge gained from experience and kept between sessions as note files the model can read later, separate from weight changes, works best when it is selective. Careful notes about what actually matters beat exhaustive archives, and an idle-time review process called Dreaming corrects notes that have drifted from what really happened.

The headline experiment used Pokemon Red. Claude Sonnet 3.7 played with an unrestricted memory, produced about 10,000 game steps, 413 files in 225 folders, and finished with one badge, at times reorganizing notes instead of playing. In Claude Fable 5, memory was capped at nine files in two folders plus a dashboard, and the model wrote plans, strategies and tested rules rather than on-screen observations, finishing with four badges. Instruction phrasing mattered as well: a rigid rule such as 'only record mistakes' wiped out most of the advantage for stronger models, while older models benefited from explicit guidance like 'review your notes before every response'. A Sonnet 4.6 chart put the score at 0.330 without memory, 0.364 with memory, and 0.641 with memory plus refined instructions.

Dreaming targets note rot. In idle periods Claude reviews transcripts of earlier sessions and checks its notes against them, building on research from Letta and UC Berkeley. In one case a faulty note put Claude on the wrong side of a mountain, where it hit the same trap again and again and missed the goal โ€” after a Dreaming pass it fixed the note and found the right path. With memory plus Dreaming Claude reached Cerulean City after 1,935 actions, while runs without it never did. About 350,000 internal documents became a linked wiki where Dreaming generates pages, fixes gaps, duplicates and broken links, and the best results came from wiki plus search. A legal test scored 87 percent with the wiki versus 73 percent with search alone on multi-document answers, with simple lookups roughly on par.

Open questions remain: mid-task note-taking distracts the agent, Dreaming costs extra compute and arrives with a delay, wikis need upkeep, and auto-loading too many memories eats context and propagates bad notes. Dreaming shipped in May as a research preview for Claude Managed Agents โ€” it reads a memory store plus up to 100 past sessions, removes duplicates and builds a reorganized store, original untouched, with Opus 4.7 and Sonnet 4.6 supported at launch. OpenAI's June memory overhaul for ChatGPT also uses the name Dreaming, with temporal accuracy moving from 52.2 to 75.1 percent per OpenAI. The comparison covers a single playthrough per model, so results are illustrative rather than definitive, and all benchmark numbers come from vendor-run tests.

Sources:
the-decoder.com/anthropic-says-the-key-to-agent-memory-is-knowing-what-to-forget-not-how-much-to-store/
www.anthropic.com/research/components-of-agent-memory
techspherenews.com/anthropic-agent-memory-quality-hinges-on-selective-forgetting
news.sulat.com/2026/11/07/agent-memory-selective-consolidation/
genaibrief.com/research/agent-memory-files-dreaming

Key figures

IMPACT checklist


Continue the conversation: https://t.me/AIAutomationArchitecture ยท Bluesky

AI Automation

Top comments (0)