Here is a thing that happens to every AI assistant, including the ones running on serious infrastructure: a user corrects the same mistake three times. The assistant acknowledges each correction, sometimes apologises, and then makes the same mistake again in the next session. The corrections do not accumulate. The assistant does not remember.

This is not a capability problem. It is a memory problem. And it is a hard one.

Self Improving Agent by @xiucheng is a skill that tries to solve it. It captures errors, user corrections, and best practices and converts them into persistent long-term memory. The goal is straightforward: make an AI assistant that learns from its mistakes rather than accumulating an ever-larger log of apologies.

Why This Problem Is Harder Than It Looks

Most people assume the fix is simple — just write corrections to a file. But that misses the actual difficulty. The problem is not storage. The problem is retrieval and relevance.

A correction logged to a flat file does nothing unless it surfaces at the right moment. If you correct an assistant for misformatting invoice emails, but the next session is about calendar management, that correction stays dormant. The assistant has no way to know these two contexts are related. What you actually need is structured recall — corrections that activate when the context is similar, not just when the exact same words appear.

This is the distinction between a scratchpad and a learning system.

How the Skill Works

The skill operates in three phases that map to how actual learning works: capture, review, and application.

Capture — When the user corrects the assistant or when the assistant identifies its own error, the skill logs it in a structured format. Not just “you did this wrong” but the error type, the context, the correction, and the principle behind it. The difference matters. “Don’t use Zapier for this” is less useful than “Zapier’s webhook timeout is too short for payloads over 50KB — use Make instead.”

Review — Periodically, the skill surfaces past corrections for review. This is where most self-improvement systems fall down — they log endlessly but never revisit. The review mechanism forces a confrontation with past mistakes at a moment when the emotional heat of the original error has cooled and rational analysis is possible.

Application — Before starting a new task, the skill checks relevant corrections and patterns from previous sessions. The goal is context-sensitive recall, not just keyword matching.

What Makes It Interesting

The most useful version of this skill is not a passive log — it is an active audit partner. The skill runs a periodic self-audit: scanning recent conversation history for repeated error patterns, flagging corrections that were applied inconsistently, and identifying cases where the same root cause produced different symptoms.

This is the part that most people do not do manually, even when they mean to. Reviewing your own mistakes is uncomfortable. An automated system that surfaces them dispassionately removes some of that friction.

The Deeper Problem It Points At

Self-improving AI assistants are a genuine research area, and for good reason. The standard approach — fine-tuning on mistakes — is expensive, slow, and risks overfitting. The prompt-injection approach — writing instructions into the system prompt — works until the prompt becomes so long it degrades performance on unrelated tasks.

What this skill represents is a middle path: structured, retrievable, context-sensitive memory of specific corrections. It is not a general solution to the learning problem. But it is a practical one for the specific, recurring mistake pattern that every long-term AI assistant user eventually notices.

Who This Is For

If you have been working with an AI assistant for more than a few weeks and have noticed the same category of mistake appearing more than once, this skill is for you. The manual alternative — keeping a corrections document and referencing it — works until the document grows beyond what you can reasonably search. The skill automates that search.

It is especially useful for agents doing complex, multi-step work where errors compound. Code review, financial analysis, content production — these are all domains where an early mistake propagates through the rest of the work and a correction at step three does not fix the output from step one.

Find it on ClawHub under self-improving-agent.