Meta AI RAM Team Introduces Self-Improving Memory for Continual Learning Agents

Meta AI's RAM team introduced Self-Improving Memory, a two-stage method that lets a model discover its own memory prompts and then trains those skills into its weights. On three Continual Learning Bench tasks, optimized prompts averaged 56.0 versus 44.6 for the best hand-written prompts, and after LoRA fine-tuning the model scored 43.1 with generic prompts versus 37.6 for training on generic runs.
Event version 1.
Key insights
No single hand-written memory strategy wins every task
A pilot study compared four human-designed memory strategies across the Database, Codebase and Spectrum tasks. No single strategy performed best on all three, motivating the model to discover its own memory design.
Harness optimization rewrites LOAD and OFFLOAD prompts
A meta agent reads task trajectories, scores and final memory directories to find where memory use falls short, then rewrites the LOAD and OFFLOAD prompts. New prompts are evaluated on a held-out set and only higher-scoring versions are kept for further revision.
Harness distillation moves part of the prompt gains into weights
The model runs with optimized prompts, the prompts are swapped back to generic ones while responses are kept, and the model is LoRA fine-tuned on the result. With generic prompts at test time, the distilled model scores 43.1 versus 37.6 for SFT on its own generic runs.
Optimized prompts beat hand-written and generic prompts
Automatically optimized prompts averaged 56.0 across the three tasks, 11.4 points above the best hand-written prompts and 19.9 points above generic prompts. The largest gain occurred on Codebase.
Search relies on evaluation to filter proposals
Of 44 revisions in completed searches, 37 failed to exceed the previous best score. In the Codebase case, the meta agent added instructions to save repository paths and failure rules after observing an ever-growing summary file.