Anthropic maps four layers of agent memory and an offline “dreaming” pass

Anthropic on 6 November 2026 published “The components of agent memory,” an engineering report by Lance Martin on how Claude agents retain experience between sessions — distinct from model weights, in-context learning and compaction — and how an offline pass it calls dreaming can reorganize that memory.

Four layers and dreaming

The design separates searchable transcripts, agent-written notes, organized representations such as linked wikis, and a small injected-context layer loaded at the start of a session. Dreaming uses idle compute to test notes against original transcripts, create or prune notes, build representations and promote a compact index into future sessions. Claude Code’s auto memory, the post says, loads only the first 200 lines or 25 KB of MEMORY.md as an index.

In a Pokémon Red test of what to save, Claude Sonnet 3.7 wrote 413 memory files across 225 folders and earned one badge, while Claude Fable 5 wrote nine files in two folders plus a top-level global.md and earned four badges (compared around step 10,000). In one navigation error, an uncorrected note about a PokéCenter’s location left Claude looping at Mt. Moon; after dreaming corrected the note against the transcript, Claude re-entered the cave and reached Cerulean City.

On an internal corpus of about 350,000 files (roughly 420 million tokens) from Stainless, a company Anthropic acquired in May, 25 questions were graded. As dreaming scaled from two sessions to 128, wiki-only performance grew from 13% to 48% of graded criteria. Search plus a growing wiki rose from 54% to 61%, versus 53% for search alone, and saved about 20,000 tokens per question. On 24 merger questions that often required combining documents, consulting the wiki scored 87% versus 73% for search alone on those multi-document items. When several agents wrote contract terms to one shared page, 41 of 57 facts (72%) survived; giving each contract its own page written by a single agent recovered 57 of 57.

Limits

These are Anthropic’s own engineering evaluations on internal and task-specific tests, not an independent public benchmark. The post documents failure modes: stale notes, a wiki that grew larger than its source documents, concurrent writers dropping facts, and injected errors that can spread to every session. Dreaming itself uses extra compute and can introduce mistakes.