What problem does it solve? Accumulated session learnings pile up in a pending backlog with no safe, reviewable way to distill them into persistent memory. This Skill consolidates the backlog into a structured proposal while guaranteeing that no memory file is edited without explicit human approval. ## Core Features & Use Cases - Two-pass proposal generation: Reads all pending learnings from .multiplai/learnings/, distills them via a generate-then-critic flow, and writes a dated proposal to .multiplai/dreams/ sorting every item into memory updates or filtered-out dispositions. - Deterministic audit sections: Adds conflict resolutions, citation repairs, routing warnings, and a rules re-learned report that identifies which existing memory rules are not sticking. - Partial-run safety: Failed chunks are never marked processed, so interrupted runs resume cleanly on the next invocation with nothing lost. - Use Case: After weeks of sessions, run the dream pipeline to turn 200+ raw learning blocks into a numbered, reviewable proposal, then apply approved changes with /multiplai-context:dream-remember. ## Quick Start Ask the assistant to run the dream skill to generate a learnings proposal from the pending backlog for review.
