Meta FAIR's Agents Write Their Own Memory Prompts, Beating Human-Crafted Ones 56.0 vs 44.6
winer666 · reddit · 2026-11-21
Meta FAIR (Jason Weston's RAM team, with UW and Johns Hopkins) published a blog on 'Self-Improving Memory', a two-stage method for continual-learning agents.
Method
- Meta-optimization: the agent (Kimi-K2.6, a 1T-param MoE) rewrites its own file-based memory prompts, searching for instructions better than the human-written seed
- Harness distillation: best runs are distilled into weights via LoRA, so memory skills survive generic prompts at test time
Results (their Continual Learning Bench: SQL, codebase fixes, radio scans)
- Self-discovered prompts average 56.0 vs 44.6 for the best hand-written harness; codebase gains largest at 64.3 (+20.6)
- After distillation with generic prompts: 43.1 vs 37.6 for plain SFT; no-memory baseline is 20.4
Honest caveats: 37 of 44 generated prompt revisions failed to beat the previous best; it's a blog post, not peer-reviewed, on the authors' own benchmark.
The takeaway: agent memory design has been hand-crafted by researchers; here the model finds better memory instructions than humans and learns them — if it generalizes, 'how should an agent remember' stops being prompt engineering.
More from coding & agent
- Pro² Agent paper: explicit Problem State lifts autonomous AI R&D to 76% MLE-Bench — rohanpaul_ai · 2026-11-21
- Google & JHU's Pro² Agent hits 76% MLE-Bench medal rate by tracking bottlenecks — rohanpaul_ai · 2026-11-21
- Stripe opens machine payments to everyone, will bake it into all new integrations — jeff_weinstein · 2026-11-21
- CCORE-Web-Coder-4B preview: 4B local model builds sites as single HTML — ConfidentDinner6648 · 2026-11-21
- Standard Bank automates genAI risk assessment on AWS with LLM-as-a-judge across 300+ controls — AWS ML Blog · 2026-11-21
- Matt Pocock: ADRs are essential for coding agents — don't remove a fence before you know why it's there — mattpocockuk · 2026-11-21