Princeton and Meta researchers teach AI agents to stop repeating their mistakes

AI agents have a memory problem. They can botch a task, get corrected, and then botch a nearly identical task the next day as if nothing happened.

Researchers from Princeton University and Meta’s FAIR division think they have a fix. Their new technique, called Self-Direction, converts an agent’s past mistakes into small, reusable skill modules, and the early benchmark results are hard to ignore.

On tested benchmarks, success rates climbed by as much as 17.0 percentage points. The trick is that the underlying model never gets retrained.

How Self-Direction works

Self-Direction gives AI agents that binder. When an agent makes a mistake and receives a corrective example, the system distills that lesson into a lightweight entry in a growing library of skills.

Each entry is a Low-Rank Adaptation, or LoRA, adapter. LoRA is a method for tweaking a large model’s behavior by training a small add-on layer instead of rewriting the whole network.

The base model stays frozen, and the adapter acts like a clip-on lens that sharpens its focus for a specific type of job. The researchers used rank-16 adapters, each touching roughly 0.02% of the base model’s parameters.

That size matters. Full model retraining is expensive and can erode capabilities the model already had, a problem the researchers designed around by keeping the base model’s original capacity intact.

The numbers

The team ran its experiments on Qwen3.6-27B models across three agent benchmarks. All three moved meaningfully in the right direction.

On GAIA, success rates jumped from 38.8% to 55.8%, a gain of 17.0 percentage points. Terminal-Bench 2.1 rose from 51.7% to 65.5%, up 13.8 points.

Claw-Eval posted the smallest gain, climbing from 64.4% to 73.3% for an 8.9-point improvement. Smallest is relative here, since that benchmark started from the highest baseline.

The library approach also held up when multiple skills were stored together. A library of three adapters lifted GAIA success from 36.4% to 52.7%.

Perhaps the most telling statistic involves repeat errors. Self-Direction raised the rate at which agents resolved previously encountered mistakes across tasks from 9.7% to 18.9%, nearly doubling it.

Crucially, that improvement didn’t come at the expense of first-time fixes. The rate of corrections on first encounters held at 67.3%.

The researchers also checked whether the adapters would step on each other’s toes. At a 0.55 similarity threshold, each benchmark’s adapter did not misroute tasks to the others, meaning the system pulled the right skill for the right job.

Why this matters for AI agents

The research was led by Yuval Kansal, an intern from Princeton, working alongside researchers at Meta. The paper was submitted to arXiv on October 22, 2026, and the project is available at self-direction.github.io.

The researchers point to research and coding environments as likely beneficiaries.

What this means

The most practical implication is cost. If teams can patch an agent’s weak spots with adapters that touch about 0.02% of a model’s parameters, they avoid the heavy compute bill and risk of full retraining.

The modular design also offers a kind of auditability. Because each lesson lives in its own adapter, a team could, in principle, inspect, update, or remove a specific skill without disturbing the rest of the system.

The routing result is worth watching closely. A skill library is only useful if the agent picks the right skill, and the clean separation at the 0.55 similarity threshold suggests the approach can scale beyond a handful of adapters without tasks getting crossed.

The repeat-mistake figure also cuts both ways. Doubling the resolution rate to 18.9% is a real improvement, but it means agents still fail to fix most previously seen errors across tasks.