Meta’s Self-Taught Memory Moved 35% of Its Gain Into Weights

Researchers at Meta AI let a model write its own rules for using memory and then trained those rules into its weights, according to a blog post that one of its authors shared on X on Tuesday. On three Continual Learning Bench tasks, the rules the model wrote scored 56.0, against 44.6 for the best hand-written ones and 36.1 for a generic setup.

Our number: after fine-tuning, the model kept 7.0 of those 19.9 points while running on the generic prompts, about 35% of the gain, by our arithmetic. The post calls that 7.0-point comparison descriptive, because the base model was evaluated separately.

Business Pill · Distillation Explained

An evergreen Business Engineer explainer on how a model learns from another model’s outputs. It teaches the general idea only and does not discuss this post.

The key insight: In this work, the model’s own prompt search added 19.9 points over generic prompts, and fine-tuning on those runs kept 7.0 of them with generic prompts at test, about 35% by our arithmetic.

What Meta Built

The setup gives an AI agent a folder of files as its only memory. Each subtask starts with an empty context, and two prompts tell the agent what to save after a task (OFFLOAD) and what to read before the next one (LOAD).

The post says no single hand-written memory design won on every task. So a meta agent reads the task agent’s runs, finds where its memory use falls short and rewrites the two prompts. Each new version is scored, and the best ones are kept for further revision.

Both agents use the same model, Kimi-K2.6. In the second stage, the agent runs with the optimized prompts, the researchers swap the generic prompts back into the recorded runs, and they fine-tune the model on the result, using LoRA with rank 64. The post says “the model is trained on memory skills it discovered itself”.

Average test score on three Continual Learning Bench tasks (Database, Codebase, Spectrum) with Kimi-K2.6: no m
Average test score on three Continual Learning Bench tasks (Database, Codebase, Spectrum) with Kimi-K2.6: no memory 20.4; generic prompts without fine-tuning 36.1; fine-tuned on generic-prompt runs 37.6; fine-tuned on optimized-prompt runs, tested with generic prompts, 43.1; best hand-written prompts 44.6; model-written (optimized) prompts 56.0. Source: Meta AI, Self-Improving Memory blog post, Figures 4 and 5, shared on X on 17 November 2026. 7.0 of 19.9 points = 35% is our arithmetic; the post calls the 7.0-point comparison descriptive because the base model was evaluated separately.

What Was Posted on X

Jason Weston, one of the researchers behind the work, introduced Self-Improving Memory on X on 17 November. He wrote that the team let the model discover its own memory skills and then trained them into its weights, and linked the blog post this piece reports.

What the Results Show

The optimized prompts outscored the best hand-written ones on all three tasks. The largest gain came on Codebase, where the score rose from 43.7 to 64.3, a 20.6-point gain. Database rose from 47.9 to 53.8 and Spectrum from 42.1 to 49.9, according to the post’s Figure 4.

With only the generic prompts at test time, the fine-tuned model scored 43.1. Fine-tuning on the model’s own generic runs gave 37.6, and the untuned model 36.1. The 5.5-point gain over generic fine-tuning is 14.6%, the post says, across ten matched evaluation seeds.

Most revisions did not help. The post says “37 of 44 generated revisions fail to beat the best score from previous iterations”, 84% by our arithmetic. Database and Spectrum peaked in the first generation; Codebase kept improving through the fourth.

Why It Matters

Today an agent’s memory habits mostly live in prompts that people write. This work moves part of that skill into the model, so it shows up under a plain prompt. It lands in a week of debate over AI that improves AI: OpenAI rated a model its first ‘High’ self-improver, and Microsoft AI’s chief executive called AI improving AI dangerous.

Meta researchers took a related route in October, when a Princeton and Meta paper trained agents’ fixes into their weights. Jason Weston, one of the authors, introduced the new work on X with the line “We let the model discover its own memory skills, then train them into its weights”.

The Business Lever

The lever is the agent harness, the prompts and tools around a model. In October, Satya Nadella said Microsoft’s harness makes AI models swappable. Meta’s method runs the other way: it trains the memory policy into the model “so that generic prompts suffice at deployment”, the post says.

The post gives no compute or cost figures. Its search runs each candidate set of prompts five times, and 37 of the 44 revisions did not improve on the earlier best.

The Structural Read

As we read the results, most of the gain still lives in the prompts: the fine-tuned model scored 43.1 with generic prompts, against 56.0 with the optimized ones.

As we read the search figures, the method depends on scoring many candidates: 37 of 44 revisions did not improve on the earlier best, and Database and Spectrum stopped improving after the first generation.

Meta AI, Self-Improving Memory blog post, 17 November 2026

“Training carries part of the gain from better prompts into the model’s weights.”

Three Implications

FOR AGENT BUILDERS The post reports that the best memory prompts differed by task, and that no single hand-written design won on all three.

FOR MODEL TRAINING The post fine-tuned Kimi-K2.6 with LoRA on runs whose optimized prompts were swapped back to generic ones, and the model scored 43.1 against 37.6 for training on generic runs.

FOR EVALUATION The post says its optimized and hand-written scores are the best observed among evaluated candidates, picked by task score.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Agentic Harness War.

The framework says: “The surface, not the model, governs autonomy.”

As we read it, this post tests the other direction: it moves part of a harness’s gain into the model, so that generic prompts carry more of the work. The framework predates this post and does not discuss it.

What Is Not Established

This is a research blog post on three tasks from one benchmark, with one model. The post says the scores are “the best observed scores among evaluated candidates”, picked by task score, and that isolating why fine-tuning helped “requires direct behavioral analysis”.

Fine-tuning recovered a share of the prompt gain, not all of it: 43.1 against 56.0 for the optimized prompts. The 35% share compares runs the post evaluated separately, so it is an estimate. We did not contact the authors.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The Agentic Harness War

A Business Engineer analysis of how value moves from the model to the harness around it, and who is building which kind.

Read the Map of AI →

The Bottom Line

Meta AI researchers had Kimi-K2.6 rewrite its own memory prompts, lifting a three-task average to 56.0 from 44.6 for the best hand-written prompts, and then trained it on those runs. About 35% of the gain stayed in the weights under generic prompts, by our arithmetic; 37 of 44 prompt revisions failed.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. Figures and quotations are from Meta AI’s blog post “Self-Improving Memory: Meta-Optimization and Harness Distillation for Continual Learning Agents”, read in full with the bar labels of its Figures 4 and 5, and from an X post by Jason Weston, one of its authors. The 19.9-point, 7.0-point, 35% and 84% figures are our arithmetic. The results cover three tasks from one benchmark with one model. We did not contact the authors.

Sources: Chen et al., Self-Improving Memory: Meta-Optimization and Harness Distillation for Continual Learning Agents, Meta AI RAM blog, 17 Nov 2026 · @jaseweston on X, 14:01 UTC, 17 Nov 2026

Scroll to Top
FourWeekMBA