Dream Learning

“Dream Learning” is an offline experience integration: an online agent performs tasks and adds immutable evidence; a background process in idle reads an experience pack, matches new against old, merges duplicates, resolves conflicts, proposes candidate updates, runs regression. Five stages: threshold trigger → orientation to current capabilities → collection and integration → verification and approval → cleanup and indexing (obsolete is archived with source and revision history preserved). Continuous evolution ≠ unbounded growth: periodic “garbage collection” of knowledge and prompts is needed.

Related: [Two Loops of Evolution], [Continuous Agent Evolution]