Meta AI's Self-Improving Memory: agents learn memory skills into their weights
faeze_brh · x · 2026-11-18
Meta AI's RAM team introduces Self-Improving Memory, a recipe for continual learning agents that lack demonstrations of good memory behavior.
Method (two stages)
- Harness optimization: a meta-agent reads the task agent's trajectories, finds memory-use shortcomings, and rewrites the memory prompts; best-scoring versions are kept and iterated
- Harness distillation: the agent runs with optimized prompts, prompts are swapped back to generic ones while keeping responses, and the agent is fine-tuned on those runs so prompt-taught skills move into weights
Key results (Kimi-K2.6 on 3 Continual Learning Bench tasks)
- Self-discovered prompts score 56.0 vs 44.6 for the best hand-written prompts (+11.4)
- Fine-tuning on optimized-prompt runs reaches 43.1 with generic prompts vs 37.6 baseline (+5.5)
Announced by Jason Weston; blog post with details available.
More from coding & agent
- SilentGuard: harness-level invariant checks cut agent API silent-failure errors from 74% to 1.7% — Genentech · 2026-11-18
- One PowerShell cmdlet, New-AgentHarness, spins up an OpenAI-powered agent — dfinke · 2026-11-18
- Doodlestein Shares His Multi-Agent Coding Stack: agent-mail + beads + ntm — doodlestein · 2026-11-18
- The bitter lesson is coming for software engineering: let agents discover abstractions — tylerjdunn · 2026-11-18
- MCP Agent Mail: open-source shared inboxes and coordination for multiple AI coding agents — doodlestein · 2026-11-18
- FrontierCS 2: benchmark makes AI agents beat published CS research blind — arena · 2026-11-18