What problem does it solve? Agents repeatedly re-derive facts that were already discovered in earlier sessions, and repo text search runs before anyone checks whether the answer already exists in memory. This Skill enforces a search-first workflow against the AiRaccoon memory server, writes durable findings back with source paths, and measures whether search results actually get used. ## Core Features & Use Cases - Search-first workflow: Run memory_search with projectId and sessionId before web search, code search, or asking the user, escalating by result quality (decisive hit, partial hit, no hit). - Watch-on-docs ritual: Mirror a project's docs directory and .semantica/ into memory with memory_watch_add so semantic search covers project documentation. - Write discipline and promotion: Write durable facts one per entry with source paths, use workspace isolation for in-progress notes, and promote cross-project facts with memory_share. - Memory-first gate hooks: PreToolUse/PostToolUse scripts deny repo text-search tools (grep, rg, find, search_files) until the session has consulted memory_search, with a 3-strike pass-through. - Follow-through measurement: A PostToolUse hook stashes search results by correlationId and records when the agent opens a returned file within 60 seconds, feeding the search_quality table. - Use Case: Starting a session on a repo, the agent watches the docs directory, searches memory before grepping the codebase, finds a prior decision with its source path, and writes the session's new finding back for the next session. ## Quick Start Ask the agent to check the memory watch status for this project, search memory for the current question before searching the repo, and write back any durable finding with its source path.
