Anthropic and agents use 'dreaming' to curate memory

Anthropic models consolidate agent memory through a background process called dreaming, which reviews work logs against existing notes to correct mistakes and prune stale information between sessions (Confluence). This architecture moves beyond simple vector retrieval by organizing memory into a hierarchy of transcripts, selective notes, and knowledge wikis rather than a single searchable attic (Shilpa Mitra from WebAfterAI). When Claude played Pokémon Red, the agent accumulated 413 memory files across 225 folders, eventually stalling due to the overhead of reorganizing its own state; a dreaming pass corrected a single erroneous map note that had left the agent trapped in a mountain cave.

Developers can implement similar routines by directing agents to update project-specific files like CLAUDE.md or AGENTS.md based on recent conversations, effectively performing a manual dreaming pass. For more complex systems, researchers at UC Berkeley developed MemGPT, which treats memory as an operating system by providing agents with a small, always-visible 'main memory' and larger external storage for paging information (Technically). To maintain reliability in production, engineers should combine these fast compaction methods with immutable logs and policy rules that preserve source evidence for high-risk actions (Rosario Fortugno from Applying AI).