Fixed memory rules hurt strong agents: "record only failures" costs Opus 5 gains
curious_vii · x · 2026-11-09
CL-Bench team shares empirical findings on agent memory engineering: simply telling Claude that a memory folder exists (what memory is for, not how to use it) was enough for Opus 4.7 and Opus 5 to use memory effectively, with noticeable performance boosts. Testing specific one-rule-at-a-time guidance, they found a "record only failures" rule stripped most of the memory benefit from stronger models — Opus 5 lost 0.234 and Opus 4.7 0.150 of the gain from memory, while weaker Sonnet 4.6 barely benefited. Takeaway: letting strong models judge salience freely beats fixed rules, which can weaken performance.
More from coding & agent
- Stanford's CS146S opens its AI-native engineering syllabus as coding agents finish homework in one prompt — Latent Space · 2026-11-09
- Telling your coding agent 'go refactor while I grab coffee' is the new workflow meme — shakoistsLog · 2026-11-09
- 20 computer-use AI agents compared: open frameworks vs managed services in 2026 — TheTuringPost · 2026-11-09
- Dev says Opus 4.5's task delegation and parallel scheduling may beat his own — lucasmeijer · 2026-11-09
- Debloating a Hisense TV with Claude Code over ADB — no root, 14 apps disabled — SeanOliver · 2026-11-09
- Connecting Grok Bot to Claude Code creates an AI loop that red-teams business plans — adamamcbride · 2026-11-09