self-improvement

To log agent errors and corrections to a structured learning log, you record non-obvious learnings and failures into templates like LEARNINGS.md and ERRORS.md within a dedicated directory. This enforces consistent frontmatter to enable future fixes across sessions.

Capturing non-obvious learnings during an agent workflow is best handled by triggering hooks after task completion or API failures. This automatically prompts logging into a structured learning log, ensuring recurring patterns are documented for continuous improvement.

Yes, you can promote broadly useful learnings to project memory in CLAUDE.md or AGENTS.md. The logging mechanism elevates validated corrections from the local learning log into agent workflows, ensuring persistent context across sessions.

Yes, this logging approach works with OpenClaw integration and cross-session prompts. It provides specific hooks and templates for OpenClaw, allowing you to apply structured logging across in-session interactions and recurring agent patterns.

To seed initial templates for tracking feature requests and knowledge gaps, you create a designated directory and populate it with standard files like LEARNINGS.md, ERRORS.md, and FEATURE_REQUESTS.md. This structures your logging immediately for continuous improvement.

You should not use a structured learning log for obvious or expected API failures. The system targets non-obvious learnings and complex corrections, so logging trivial errors clutters the project memory and reduces the effectiveness of cross-session integration.