Meta FAIR researchers have released a method called Self-Improving Memory. The technique allows AI agents to write their own instructions for managing file-based memory. A pilot study found that the AI-generated prompts achieved an average score of 56.0 on the Continual Learning Bench, beating the best human-written strategies which scored 44.6.
In this article
The core setup
The project uses the base model Kimi-K2.6. This is a 1T-parameter MoE with 32B active parameters and a 256K context window. The team added a LoRA adapter with a rank of 64 to the attention projection layers. Vast.ai suggests an 8x H200 node for serving Kimi-K2.6, though native INT4 weights are available.
The method relies on two prompts: LOAD and OFFLOAD. LOAD retrieves notes before a subtask begins. OFFLOAD updates the memory directory after the task. The system starts with four human-written seeds: Summary, Retrieval, Skill, and Hypothesis. The meta agent rewrites these prompts by running the full task stream five times for each candidate. The top four results by mean score become parents for the next revision cycle.
How distillation works
Stage two follows a Discover-Demonstrate-Distill pipeline. The task agent runs subtasks using the optimized prompts. Every model call is logged. The research team then swaps the optimized prompts back to generic ones in that history. The model is fine-tuned with LoRA rank 64 on attention projection layers. The loss covers only assistant tokens, including reasoning and tool calls.
The meta agent and task agent are the same Kimi-K2.6 model. The model is trained on memory skills it discovered itself. This 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.
Performance results
The tests use three tasks from the Continual Learning Bench: Database (SQL), Codebase (repo fixes) and Spectrum (radio scans). Test subtasks follow a designed phase shift. This measures generalization to new instances of the same tasks.
Score breakdown
The table below shows the test scores. Kimi-K2.6 is used for all rows.
| Test score | Database | Codebase | Spectrum | Average |
|---|---|---|---|---|
| No memory | 7.5 | 31.8 | 21.9 | 20.4 |
| Compaction | 39.8 | 36.9 | 34.6 | 37.1 |
| Generic harness | 32.3 | 33.5 | 42.5 | 36.1 |
| Best human-written | 47.9 | 43.7 | 42.1 | 44.6 |
| Optimized harness | 53.8 | 64.3 | 49.9 | 56.0 |
| SFT on generic trajectories | 31.6 | 37.5 | 43.6 | 37.6 |
| SFT on optimized trajectories (generic prompts) | 37.7 | 45.1 | 46.5 | 43.1 |
Source: Figures 4 and 5, Self-Improving Memory blog.
Optimized harnesses beat the best hand-written seed on all three 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%.
Comparison with other methods
Several recent methods search over agent memory or harnesses. Benchmarks differ, so headline numbers are not directly comparable.
| Feature | Self-Improving Memory | ALMA | MemEvolve | Harness-Zero |
|---|---|---|---|---|
| Org | FAIR at Meta, UW, JHU | Xiong, Hu, Clune | OPPO AI Agent Team, LV-NUS | Peking University, Google, HKUST |
| What is searched | LOAD/OFFLOAD memory prompts | Memory designs as Python code | Memory architecture and contents | Full agent harness |
| Trains weights | Yes, LoRA rank 64 | No | No | Yes, LoRA SFT |
| Models | Kimi-K2.6 (agent and meta agent) | GPT-5 meta agent, GPT-5-nano/mini task agent | GPT-5-mini (plus DeepSeek V3.2, Kimi K2) | Qwen3.5-9B student |
| Benchmarks | Continual Learning Bench (3 tasks) | ALFWorld, TextWorld, Baba Is AI, MiniHack | GAIA, WebWalkerQA, xBench-DeepSearch, TaskCraft | SpreadsheetBench Verified, AppWorld, USPTO Retrosynthesis |
| Headline result | 56.0 vs 44.6 best human harness; 43.1 distilled | 53.9% vs 41.1% no memory (GPT-5-mini) | Up to 17.06% gains | 23.3% to 44.3% with harness removed |
| Code | Not released | GitHub | GitHub | GitHub |
Harness-Zero is the closest match, but it uses a separate agent to correct the student. Self-Improving Memory skips that step.
What it means
The research shows that an AI can improve its own workflow by rewriting its memory instructions. This is an internal research development. It does not imply changes to customer products, accounts, privacy or API behaviour. The team has not released the code.


