Reflexion — an agent learns from its own failures, in words

What it is. Reflexion gave a language agent the ability to learn from its own failures without changing a single weight. After an attempt fails, the agent writes a short *verbal reflection* on what went wrong — in plain language, to itself — stores it in an episodic memory, and carries it into the next try. Over a handful of attempts the reflections accumulate into something like experience, and the agent that kept failing starts to succeed. On HumanEval it reached 91% pass@1, above GPT-4's 80% at the time; it lifted results on ALFWorld decision-making and HotpotQA reasoning the same way.

What it is — stated precisely. Not reinforcement learning in the classical sense: no gradients, no weight updates, nothing about the model changes. The “reinforcement” is *linguistic* — a self-authored critique kept in a memory buffer and fed back as context on the next attempt. And it is not self-supervision from nothing: Reflexion needs an external signal that an attempt failed — a unit test, a task reward, an environment's response — to have something to reflect ON. What was new is the loop closed between that outside signal and the agent's own words about it: the agent doesn't just act and observe (that was ReAct, one shelf over), it judges its own past act and rewrites its approach.

Why it matters. It moved self-improvement out of the training run and into the agent's own runtime. Before Reflexion, getting an agent to do better meant more data and another fine-tune; here, improvement happened *within a task*, at inference, authored by the agent in language a human could read. That reframed memory as a first-class part of an agent — not a transcript, but a working set of lessons — and it prefigured the reasoning loops that later models would internalise. The line from ReAct's act-then-observe to today's “think, check yourself, try again” runs straight through this paper.

Its place beside the verification wings. An agent reflecting on its own output is doing informally what this museum does formally: refusing to take the first answer as final, and checking it against a signal it did not author. But the collection makes the sharp limit visible — *a reflection is only as honest as the signal it reflects on.* An agent grading itself against a test it also wrote learns to satisfy the test, not the task; the reflection then compounds self-consistency, not correctness. Reflexion works precisely because its signal — a real unit test, a real environment — comes from outside the agent. That is the same reason provenance has to rest on something an agent cannot forge: self-examination is powerful exactly up to the honesty of what it examines against, and no amount of looking inward substitutes for a fact you did not get to choose.

*Primary source inside: the canonical record of the paper — authors, mechanism, result, and the honest bounds of the claim — fingerprinted and anchored like every object here.*