Anthropic: agent memory works best by forgetting selectively—notes beat 400+ archives 4x in Pokémon Red
The Decoder · rss · 2026-11-16
Anthropic studied how Claude retains knowledge across sessions and found selective notes beat exhaustive archives by a wide margin.
- In a Pokémon Red test, the newer Claude Fable 5 progressed four times further than Sonnet 3.7 while keeping just a handful of strategy files, versus Sonnet 3.7's 400+.
- A "Dreaming" process retroactively corrects faulty memories.
The key insight: the bottleneck for agent memory isn't how much you store, but knowing what to forget.
More from coding & agent
- User's $300 subscription generated $7,492 of usage — 25x value on Codex and Claude — theteknosaur · 2026-11-16
- Your agent's inference endpoint sees all its memory: treat it as a data boundary — krishnan · 2026-11-16
- Dev ships a gladiator management game built entirely with Claude Code — StarDrifterG · 2026-11-16
- Rust rewrite of mcbench-lite cuts first start from ~71s to under 1s with AI coding — adonis_singh · 2026-11-16
- Solo dev ships his son's game idea to iOS and Android for under $300 — morganinuk2 · 2026-11-16
- Enterprise AI coding fails on people, docs and process — not the codebase — _jaydeepkarale · 2026-11-16