Meta FAIR's self-optimized AI memory scores 56.0 across three tasks

Executive Briefing
“Meta FAIR has introduced a method that lets an AI agent discover how to store and retrieve...”
Key Takeaways
- 1.The method uses file-based memory. One instruction updates it after a subtask, while another retrieves notes before the next subtask.
- 2.During the search stage, an agent rewrites the two memory instructions. Each candidate runs through the full task sequence five times, and selection uses a holdout set the agent cannot see.
- 3.According to the article, the optimized instructions averaged 56.0 across Database, Codebase and Spectrum, versus 44.6 for the best human-written option. All reported result rows used Kimi-K2.6.
- 4.The team also used runs with optimized instructions to fine-tune the model. When both versions used generic memory instructions, this distilled version averaged 43.1 across 10 matched seeds, versus 37.6 for fine-tuning on runs made with generic instructions.
- 5.The search did not improve consistently with each revision. The article reports that 37 of 44 revisions failed to beat the previous best.
Meta FAIR has introduced a method that lets an AI agent discover how to store and retrieve information between subtasks. In the results reported by the article, the discovered instructions averaged 56.0 across three tasks, compared with 44.6 for the best human-written instructions.
Key Developments & Primary Facts
The method uses file-based memory. One instruction updates it after a subtask, while another retrieves notes before the next subtask.
During the search stage, an agent rewrites the two memory instructions. Each candidate runs through the full task sequence five times, and selection uses a holdout set the agent cannot see.
According to the article, the optimized instructions averaged 56.0 across Database, Codebase and Spectrum, versus 44.6 for the best human-written option. All reported result rows used Kimi-K2.6.
The team also used runs with optimized instructions to fine-tune the model. When both versions used generic memory instructions, this distilled version averaged 43.1 across 10 matched seeds, versus 37.6 for fine-tuning on runs made with generic instructions.
The search did not improve consistently with each revision. The article reports that 37 of 44 revisions failed to beat the previous best.
Strategic Analysis & Implications
The 56.0 score measures the optimized instructions, not the fine-tuned model. The 43.1 versus 37.6 result suggests fine-tuning retained some benefit after a return to generic instructions, but did not reproduce the optimized instructions' full score.
Because many revisions brought no improvement, testing and selection are central to the method, rather than simply having an agent write more instructions.
Outlook & Factors to Watch
The article does not disclose the training hardware, leaving the cost of reproducing the fine-tuning step unclear.
The tests use new instances of the same three tasks. They do not establish how the method would perform on different kinds of work.







