What changed

In a November 6 engineering report, Anthropic describes memory that persists between agent sessions: transcripts, notes, linked representations and selected context loaded into a new session. It calls offline work to check or reorganize these records dreaming. The report distinguishes this stored context from changing model weights and from within-conversation compaction.

Source: Anthropic: The components of agent memory

Our analysis

Our analysis: an agent team can start with a small index of durable, source-linked notes and retain a way to inspect original records. Evaluate whether retrieval and periodic corrections help on representative tasks before committing to a larger automatically generated knowledge base.

Keep in mind

Anthropic reports task-specific experiments, not a universal reliability guarantee. In its database benchmark, an instruction to record only failures reduced memory gains for the two Opus models tested. Offline review also consumes compute, and stored notes can be inaccurate or become outdated. We have not reproduced these tests.

Source: Anthropic: The components of agent memory

Sources & dates

Article updated: 8 Nov 2026

Editorial standards & corrections