How AI Agents Remember and Reuse Skills (Without Breaking)

An agent that solves a task should get smarter for the next one. In principle, it can induce a skill from what it just did — "when I need to do X, do Y" — and reuse it later. In practice, this is where most agent memory systems quietly fall apart.

A 2026 paper (Break It Down, Pass It On) ran a controlled study on when skills actually transfer. The results are a practical manual for anyone building agent memory.

The two mistakes everyone makes

The researchers compared a few approaches and found two traps:

Trap 1: Task-level skills hurt. Skills captured at the whole-task level (e.g., "how to do a full data-analysis task") mostly performed worse than having no memory at all. They were too vague to apply and too rigid to adapt.

Trap 2: Code skills transfer worse than text. Skills written as code snippets underperformed skills written as plain-language descriptions. Text is more flexible; code locks in an implementation that rarely matches the next task.

What actually works

Three findings you can use today:

1. Decompose to the subtask level. Skills like "how to clean a malformed CSV column" transfer far better than "how to analyze sales data." Granularity is everything.
2. Prefer text over code. A natural-language description of a skill generalizes across tasks that code can't.
3. Score skills before you trust them. The paper introduces a skill utility score that combines specificity (how closely the skill matches real tasks) and abstractness (how evenly its usefulness spreads across tasks). Crucially, you can compute it from the skill and task descriptions alone — no execution required.

Why the utility score is the real gem

Most memory systems dump skills into a library and hope. The utility score gives you a cheap pre-flight check: before an agent deploys a skill to a new task, score it. Low score? Skip it. This single idea prevents the "task-level skills make things worse" failure mode.

Why this matters for agents

Memory is the component agents most often break on in production. Context windows fill up, relevant details get dropped, and the agent starts hallucinating its own task state. A disciplined, subtask-level, text-based skill memory — gated by a utility score — is one of the few evidence-backed ways to make agent memory actually help instead of hurt.

This is the memory system component in the standard agent anatomy — the context that keeps an agent on track across many steps. How memory fits into the full perceive-reason-act loop is documented step by step in the 2026 AI agent guide. See also AI agent memory management.

Related: The AI agent architecture · AI agent guide · What is an AI agent? · AI agent memory management

how do AI agents work — return to the complete AI agent architecture guide.