Meta FAIR Self-Improving Memory Hits 56.0 on Continual Learning Bench

Meta FAIR has introduced Self-Improving Memory, a method for AI agents that manage their own memory. The model first discovers its own memory instructions. It then trains those skills into its weights. Co-authors include researchers from the University of Washington and Johns Hopkins University.

TL;DR

  • Method: The base is Kimi-K2.6, a 1T-parameter MoE with 32B active and 256K context. Distillation adds a LoRA adapter (rank 64) on attention projections.
  • Runs on: Training hardware is not disclosed. Kimi-K2.6 serving needs a multi-GPU node; Vast.ai recommends 8x H200. Native INT4 weights exist.
  • Performance: Self-found memory prompts average 56.0 on 3 tasks, versus 44.6 for the best human-written ones.
  • Best: Codebase, 64.3 with optimized prompts, +20.6 over the best hand-written harness.
  • Worst: Spectrum distillation adds only 2.9 points (46.5 vs 43.6 for SFT on generic runs).

What is Self-Improving Memory?

Self-Improving Memory is a 2-stage method from Meta FAIR that teaches an agent model to manage file-based memory. The agent works through a stream of subtasks. Each subtask starts with an empty context. Only a persistent memory directory, edited with file tools, carries over.

Two prompts control that memory. OFFLOAD updates memory after a subtask. LOAD retrieves notes before the next one. Generic versions leave the strategy to the model. A pilot study in the research post found that no single hand-written strategy wins every task.

How does harness optimization work?

In stage 1, a meta agent rewrites the LOAD and OFFLOAD prompts. The search starts from 4 human-written seeds: Summary, Retrieval, Skill and Hypothesis. Each candidate runs the full task stream 5 times. The top 4 by mean score become parents for revision.

For each parent, the meta agent reads trajectories, scores and final memory directories. It then proposes revised prompts. Selection uses a holdout set the meta agent cannot see.

The discovered prompts are concrete. One Database prompt says to read GUARDRAILS.md, then SCHEMA.md, before answering. The best Codebase design uses a quickstart file, an indexed fix registry and separate fix files.

How does harness distillation move the gains into weights?

Stage 2 follows a Discover-Demonstrate-Distill pipeline. The task agent runs train subtasks with the optimized prompts. Every model call is logged. The research team then swaps the optimized prompts back to the generic ones in that history.

The swap works because the behavior stays valid under generic instructions. Examples include keeping QUICKSTART.md, REGISTRY.md and SCHEMA.md files, and verifying join keys before aggregation. The model is fine-tuned with LoRA, rank 64, on attention projection layers. The loss covers only assistant tokens, including reasoning and tool calls. Each task gets equal weight.

The meta agent and task agent are the same Kimi-K2.6 model. So the model is trained on memory skills it discovered itself.

What do the benchmarks show?

The tests use 3 Continual Learning Bench tasks: Database (SQL), Codebase (repo fixes) and Spectrum (radio scans). Test subtasks follow a designed phase shift. So results measure generalization to new instances of the same tasks.

Test scoreDatabaseCodebaseSpectrumAverage
No memory7.531.821.920.4
Compaction39.836.934.637.1
Generic harness32.333.542.536.1
Best human-written47.943.742.144.6
Optimized harness53.864.349.956.0
SFT on generic trajectories31.637.543.637.6
SFT on optimized trajectories (generic prompts)37.745.146.543.1

Source: Figures 4 and 5, Self-Improving Memory blog. Kimi-K2.6 for all rows.

Optimized harnesses beat the best hand-written seed on all 3 tasks. Codebase gained the most, at 20.6 points. Search was noisy: 37 of 44 generated revisions failed to beat the previous best. Database and Spectrum peaked in the first generation. Codebase kept improving through generation 4.

After distillation, every model runs the generic harness. The distilled model averages 43.1 across 10 matched seeds. SFT on generic trajectories reaches 37.6. That is a 5.5-point gain, or 14.6%.

Source link

Leave a Reply

Your email address will not be published. Required fields are marked *