self-correction-loop

To stop repeating mistakes, you need a self-correction loop that captures explicit user corrections, generalizes them into reusable project rules, and stores them in a memory file like MEMORY.md. This ensures past feedback guides future outputs automatically.

Rule generalization works by taking a specific correction, such as requesting a different cache implementation, abstracting it into a broad principle, and categorizing it in project memory. This prevents memory bloat and creates compounding guidance for future tasks.

The best way to manage project conventions is to use a self-correction loop that scans existing project memory to deduplicate overlapping rules via overlap checking. It updates current entries instead of adding duplicates, preventing memory bloat and maintaining clean conventions.

To apply stored project rules at the start of a new AI session, use a self-correction loop mechanism that instructs the assistant to review existing rules in MEMORY.md immediately. This ensures captured guidance and generalized conventions are applied before generating new outputs.

Yes, you can use a self-correction loop to update MEMORY.md automatically. After detecting a correction, the process handles rule generalization, deduplication, categorized storage, and confirms exactly what was added or updated, removing the need for manual memory formatting.

Without a memory capture mechanism, correcting an AI output remains a one-off fix. The assistant will likely repeat the same mistake in future sessions because the specific feedback is not generalized into a reusable project rule or stored for future review.