continuous-learning-agent

Create a persistent learnings directory such as .Codex/learnings/ with subfolders for patterns, mistakes, decisions, and context. After each task, log structured Markdown entries capturing root causes, fixes, and reusable patterns so future sessions can retrieve them.

Use error log entries recording date, context, root cause, fix, and prevention rules, plus success pattern entries capturing reusable approaches and code templates. Store them in categorized files under learnings/patterns and learnings/mistakes for later grep-based retrieval.

Yes. A post-task hook appends the task name, approach, outcome, and learning to a dated daily log file, while a pre-task hook greps past learnings for similar tasks and known pitfalls before work begins.

Use grep recursively across the learnings directory, for example grep -r "pagination" .Codex/learnings/patterns/ to find similar solutions, or grep for "Lessons Learned" headers to extract synthesized insights from all entries.

File-based memory relies on keyword grep matching, so retrieval quality depends on consistent terminology in log entries. It also requires disciplined manual or hook-driven logging; skipped entries create gaps that future sessions cannot recover.