Anthropic’s Lance Martin describes memory for AI agents as a stack: searchable transcripts, notes, linked reference pages and a small index loaded into each new session. Background passes can check notes against the original records and reorganize them. Anthropic calls this process “dreaming.”
- In one test, a wiki plus search met 61% of graded criteria across 25 questions about company documents. Search alone met 53%; the wiki alone reached 48%. The wiki took 128 preparation sessions to build.
- A legal-document example lost facts when agents wrote to the same table. After the team tuned instructions to give each contract its own page and quote terms directly, all 57 graded facts survived, up from 41.
- A wiki offered no clear gain on one task where answers were easy to find in individual documents. Another grew larger than its source material.
These are Anthropic’s own task-specific tests, not a general guarantee that background memory maintenance pays for itself. The morning AI News edition also covered the report.
The 81-comment thread on Hacker News adds practical reservations and approaches:
What the thread adds
- yfontana — says most automatically created memories concern a single task and describes analyzing agent transcripts for recurring tool failures and efficiency problems.
- prtmnth — describes a pruning skill that routes a memory to code, project instructions, global instructions or deletion.
- benfortuna — argues that memory systems need provenance so earlier records survive conflict resolution.
- andai — questions whether the extra gains from a growing wiki justify the cost of background processing, compared with search.
HN handles are pseudonymous; HN publishes no per-comment scores. This is a slice in HN’s ranking order, not a consensus.