Continuous Learning Kit

Problem being solved

Large models often hit the same wall in multi-session agent work: context is easy to lose, making project details, user preferences, and failure causes hard to retain across days; task boundaries are not explicit enough, so switching from LIMS queries to report generation can mix in irrelevant information; and errors do not accumulate, causing repeated API guesses or workflow gaps.

How it works and limits

The kit separates short-term capture from long-term analysis. In the short cycle, it judges whether each message starts a new task and records key conversation data into MemPalace. In the long cycle, it runs daily sync, a 02:00 dream analysis pass, extraction, documentation updates, and self-correction through ERRORS.md and LEARNINGS.md. It turns failure patterns from conversations into reusable rules rather than keeping raw chat history only. With a configured model, it can synthesize deeper lessons; without one, it mainly does basic classification. It fits multi-project work, preference memory, and skill workflow refinement, but it depends on MemPalace, and scheduler status, database permissions, and memory size can affect stability.