On 2026-11-06 Anthropic published The components of agent memory, covering in-session memory notes plus offline dreaming (fix notes against transcripts, build wikis, promote hierarchy). It reports CL-Bench, Pokemon, and internal-corpus wiki evals; @AnthropicAI amplified. Engineering methods post, not a product GA.
Key Takeaways
- ✓Source: Anthropic Engineering The components of agent memory (2026-11-06) + @AnthropicAI
- ✓Core: memory notes in-session plus dreaming offline to fix notes, build wikis, and preload top-layer memory
- ✓Evals: on CL-Bench DB tasks Opus 5+ closes the learning loop; in Pokemon Fable 5's 9 files beat Sonnet 3.7's 413-file archive
- ✓Wiki: on ~350k internal docs, dreaming 2 to 128 sessions lifted 25-question scores ~13% to ~48%; search+wiki ~61% vs search-only ~53%
- ✓Practice: Claude Code auto memory injects only first 200 lines or 25KB of MEMORY.md; methods post, no new public API announced
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Core Background & Industry Pain Points
Long-horizon agents must keep experience across sessions, but writing notes mid-task distracts and bad notes poison later runs. Anthropic's 2026-11-06 Engineering post maps online encoding plus offline consolidation to agents: write memory notes in-session, then dream while idle to organize and correct them.
Architecture Highlights & Internals
Memory notes are retrievable cross-session knowledge (not weight updates or single-turn in-context learning). Dreaming validates notes against transcripts, builds linked wikis, promotes rules into session-start preload, and shapes memory toward a goal. Claude Code auto memory injects only the first 200 lines or 25KB of MEMORY.md so instruction files stay short enough to follow.
Authoritative Benchmarks & Measured Scores
Vendor examples only: on CL-Bench, Opus 5+ writes evidenced rules; in Pokemon, Fable 5's nine files and four badges beat Sonnet 3.7's 413-file one-badge archive. On ~350k internal docs, wiki scores rose ~13% to ~48% across dreaming sessions; search+wiki ~61% vs search-only ~53%. No third-party leaderboard.
Developer Hands-on Guide
Mount a persistent memory folder with Managed Agents or Claude Code; vague "memory folder available" guidance often beats rigid "record only failures" rules. Follow the Engineering post and Claude Code docs; this item announces no new memory API or pricing—do not conflate with Meta Active Memory product features.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.