# Anthropic says agent memory depends on forgetting and correcting notes

published: 2026-11-16T19:05:51.338Z
topic: memory
source: https://the-decoder.com/anthropic-says-the-key-to-agent-memory-is-knowing-what-to-forget-not-how-much-to-store/
url: https://aiagentnews.io/stories/anthropic-says-agent-memory-depends-on-forgetting-and-correcting-notes

Anthropic's research on Claude's cross-session notes found that selective notes beat exhaustive archives. A post-session process called Dreaming, which checks notes against past transcripts, can correct faulty memories.

Anthropic describes how Claude keeps knowledge across sessions by writing notes to a folder it can read later, while the model itself stays unchanged. In a Pokémon Red test, the older Claude Sonnet 3.7 saved nearly everything, creating 413 files in 225 folders over about 10,000 game steps, and at times spent its effort reorganizing notes instead of playing. The newer Claude Fable 5 kept its memory to a few files focused on plans, strategies and verified rules, and earned four badges against one for the older model.

Dreaming runs after a session, when Claude reviews earlier transcripts during idle periods and checks its notes against them. In the Pokémon test, a wrong note about a mountain landmark sent Claude into the same trap three times until a Dreaming pass corrected it. For larger knowledge bases, Claude built a linked wiki from about 350,000 internal documents from Stainless, and in a legal test it scored 87 percent with the wiki against 73 percent with search alone.

Anthropic says open questions remain. Taking notes mid-task distracts the agent, Dreaming adds compute cost and delivers results with a delay, and wikis need regular upkeep. Dreaming was introduced in May as a research preview for Claude Managed Agents, where it reads a memory store and up to 100 past sessions and produces a reorganized store. All of these results come from Anthropic's own tests, and the Pokémon comparison covers a single playthrough per model.

## Why it matters if you build agents

Developers building long-running agents should treat stored notes as claims that can go stale or be wrong, and verify them against the record before acting on them. Keeping notes to strategies and checked rules may work better than logging every step, and loading too many memories into each session can carry bad notes forward. Rigid rules such as "only record mistakes" reportedly reduced the gains for stronger models, so explaining what memory is for may be the better approach.
