ainb-reflect-memory

Build a long-lived, filtered SLICE of the global learnings KB and answer a question-set over it ("ask the auth subsystem", "ask the migration log"). Unlike /reflect:recall (a fresh hybrid query each time), a corpus is a saved filter snapshotted to disk — prime it once, ask many questions, reprime when the KB drifts. Use when the user wants a durable Q&A session against one filtered area of their own code history rather than a one-shot search.

Report reflect drain spend over a time window — tokens split by cached (cache_read), uncached writes (cache_creation), and io (input+output), with a $ estimate, grouped by day / outcome / model / transcript. Reads the drainer's cost log and surfaces outlier runs and cache-reuse health (the 41.5M-token failure mode = low cache reuse + high cache writes). Also reports the recall followup rate (the recall-quality self-monitor: how often a session searched again within 30s and got disjoint results). Use to answer "what is reflection costing me" for the last day / week, or "is recall satisfying".

Export or import the reflect knowledge base as a deterministic, git-friendly tarball — the learnings markdown corpus plus the reflect.db rows in one snapshot. Use to back up the KB, move it to another machine, or restore it: `export` writes the tarball, `import` restores it (with an optional GraphRAG reindex). Answers "snapshot my reflect KB", "move my learnings to another machine", or "restore my knowledge base from a backup".

The global knowledge indexer. Harvests ALL memory sources across all tools (Claude, Codex, Copilot, Gemini) and all project types into the unified GraphRAG + QMD knowledge base. Archives originals, generates entity sidecars, and dual-indexes for future retrieval. This is THE command that makes the knowledge base comprehensive.

Full conversation scan for self-improvement. Detects behavioral corrections and knowledge signals, classifies them, proposes agent updates and knowledge notes with entity sidecars for GraphRAG indexing. Correct once, never again.

Retrieve relevant prior learnings from the global knowledge base. Hybrid vector + graph search over 170+ indexed learnings, reranked by confidence, recency, and tag overlap. Use when starting work, debugging a recurring problem, or before implementing a feature that may have prior art.

Show reflection metrics, pending reviews, sidecar coverage, and GraphRAG health. Read-only views into the reflect system state. Can also approve/reject pending low-confidence items.

Project-level memory consolidation. Merges orphaned worktree memory directories into a single .agents/MEMORY.md for the current project. Deduplicates, sections, and proposes cleanup of orphan dirs. Does NOT index into the global knowledge base — use reflect:ingest for that.

Triage and acknowledge entries in the reflect errors sink (~/.reflect/errors.json). Invoked from the statusline ⚠N badge when pipeline errors accumulate (drain poison, parser crashes, ingest failures, hook timeouts).

Read and edit the pinned memory slots — a fixed set of named, size-capped, agent-editable scratchpads (persona, user_preferences, tool_guidelines, project_context, guidance, pending_items, session_patterns, self_notes). Slots are the agent's fast working memory: they auto-inject at SessionStart ahead of any recall results. Use to record durable preferences, project context, pending work, or notes-to-self mid-session.

Verify Reflect end-to-end health, especially drain, ingest/consolidate maintenance, launchd timers, statusline error surfacing, and real Claude/Codex harness smoke checks. Use when Reflect seems stuck, drain is not clearing pending reflections, ingest has gone quiet, statusline ERR needs proof, or a change needs tmux-backed validation.