self-improving

Agent memory persistence captures explicit corrections and self-reflections after task failures or rejections, storing them in a local tiered directory to prevent repeating mistakes across sessions.

Self-criticism memory records what went wrong and the lesson learned after feedback, promoting useful patterns into confirmed preferences and active workflows through a structured local index.

Yes, agent self-improvement memory is scoped across global, domain, and project levels, allowing you to restrict learned preferences and corrections to specific local contexts.

No, heartbeat maintenance uses conservative rules and heartbeat-state markers to decide when to process error recovery and update memory automatically, avoiding unnecessary memory churn.

The best way to apply user corrections is to reject the outdated output, ask the agent to proceed, and let it log the correction to update local memory for the next execution.

Local learning memory protects confirmed knowledge using conservative heartbeat rules to avoid churn, meaning it prioritizes stable preference retention over aggressive or automatic knowledge overwrites.