RL-taught memory: training models to use their notes is cheap and robust
realJessyLin · x · 2026-11-05
New research from Engramlab shows training and harness memory consistently stack — the strongest systems use all kinds of memory. Teaching models to use their notes with RL augments base models with memory, and taking gradients is surprisingly cheap. Memory use stays robust even through millions of tokens of edits. The authors predict harness/prompt engineering will matter less as models learn to operate with what they have — managing docs and skill.mds is itself becoming a skill as text memory and raw context blur.
Related event: Engram: RL Teaches Models to Use Text Memory at Surprisingly Low Cost(4 posts)→
More from AGI Musings
- If your decision theory says voting is useless, it may not be good decision theory — JeffLadish · 2026-11-05
- AI video imagines year 2126: monopoly, containment or abundance after labor — No_Twist_678 · 2026-11-05
- AI2 exec Bindu Reddy: AI's reputation is unfairly bad and the industry should celebrate it — bindureddy · 2026-11-05
- Many grant solicitations now flagged as 100% AI-generated by detector Pangram — analisereal · 2026-11-05
- Why AI research keeps looking under the lamppost: stale problem frames inherited from Kant — fkasummer · 2026-11-05
- Michael Nielsen on why scientists can't tell an anomaly from a revolution — michael_nielsen · 2026-11-05