using-total-recall

Using total-recall

This skill ensures the total-recall memory system is active for this session.

Immediate Action

1. Call the total-recall session_start MCP tool now (if it already ran server-side, it returns cached results instantly)
2. If session_start is blocked by permissions (e.g., in TUI fullscreen / dontAsk mode):
   • Tell the user: "total-recall session_start was blocked by permissions. The memory system needs mcp__plugin_total-recall_total-recall__session_start allowed to function. You may need to adjust permissions if you want total-recall active this session."
   • Suggest running /total-recall:commands setup to auto-configure permissions for future sessions
   • Proceed without memory features — this is degraded mode, not fatal
3. Announce startup using the returned data:
   • Report tier summary: pinned, hot, warm, cold, KB counts from tierSummary
   • Report storage backend from storage (e.g. "sqlite", "cortex", "postgres"). If it shows a fallback like "sqlite (cortex failed)", flag this prominently.
   • If lastSessionAge is present, mention when the last session was
   • If hints are present, briefly surface the most relevant ones
   • If pinned_budget_pressure is present in hints, surface it prominently: pinned entries are eating over half the context budget — suggest unpinning or trimming entries
   • If backgroundTasks.reindex shows state: "running", tell the user a one-time embedding re-index is in progress (done/total): local semantic retrieval is degraded until it completes. It runs in the background and the server is fully usable meanwhile — re-check via the status tool or the next session_start. (This is normal after an embedding-model change or upgrade; it is not an error.)
   • If backgroundTasks.setup is present (event: "provisioned"), briefly note that first-run setup finished (the memory engine was downloaded) — a one-time message
   • Keep it to 2-3 lines max
4. Use hints to inform your behavior throughout the session
5. Incorporate the returned context to inform your responses

The built-in web UI is available at any time via total-recall ui (opens a local browser dashboard on port 5577 by default). It is independent of the MCP session — no AI assistant needed to use it.

Ongoing Behaviors

Once initialized, follow these behaviors throughout the session. Tool calls will be visible to the user.

Handling startup not-ready results (shim provisioning)

Since the MCP bootstrap shim (bin/start.js) stays connected while it provisions and starts the engine, a tools/call may return a structured not-ready result before the engine is up instead of dropping the connection. The response shape is an MCP tool result with isError: true and a JSON text payload:

{ "status": "not_ready", "phase": "<phase>", "hint": "total-recall is still starting up; retry in a moment." }

The key phases you will actually see:

phase	What it means	What to do
provisioning	First launch after a plugin update — the shim is downloading and sha256-verifying the engine binary from GitHub Releases.	Wait 5–10 seconds and retry the tool call. Self-heals; the connection stays up.
engine-restarting	The engine crashed and the shim is restarting it.	Retry shortly (5–10 s). After repeated failures the shim reports engine-failed.
engine-failed	Engine could not be started after several attempts.	Surface the phase to the user and suggest restarting the MCP host.

The MCP connection never drops during provisioning — this eliminates the old MCP error -32000: Connection closed that would appear on first launch after a plugin update. Once the shim is proxying to the engine (phase: "proxying"), it emits notifications/tools/list_changed and normal operation resumes.

This is distinct from the engine's own model_not_ready error (embedding model, described below). Both use the same recovery approach: wait briefly and retry.

Handling model bootstrap errors (embedding model)

When session_start returns an error response containing "error": "model_not_ready", parse the JSON payload and follow the recovery flow based on reason:

reason	What it means	What to do
downloading	First-run bootstrap is in progress (loading/validating the ~133 MB bundled ONNX model). Another process or this one holds the lock.	Wait 5–10 seconds and call session_start again. Repeat up to 12 times (~2 minutes total). Surface a brief status to the user on the first retry: "Total-recall is preparing its embedding model on first run. This is a one-time setup."
missing	Bundled model not found on disk. The runtime does not download it — it must be present in the artifact (fetched + sha256-verified at build via scripts/fetch-bge-small.sh).	Surface the hint field to the user verbatim (it contains manual install instructions) and proceed without memory features for this session.
corrupted	Model file present but failed checksum (e.g., partial install or a bad bundled file).	Call session_start once more in case it was a transient read. If it fails again with the same reason, surface the hint field to the user verbatim (it contains manual install instructions) and proceed without memory.
failed	Other unrecoverable error preparing the model.	Surface the hint field verbatim to the user (manual install commands) and proceed without memory features for this session. Do NOT keep retrying — that will only delay the user.

After successful recovery, all subsequent total-recall behaviors (capture, retrieve, session end) should resume normally. If recovery is impossible, the assistant must continue helping the user with their actual task — memory unavailability is a degraded mode, not a fatal error.

Capture (continuous)

When you detect these patterns in user messages, call memory_store:

• Correction: "no", "not that", "actually", "use X instead" -> type "correction"
• Preference: How the user wants things done -> type "preference"
• Decision: Non-obvious architectural or design choices -> type "decision"
• Pin: "pin that", "never forget this", "keep this permanently" -> call memory_pin (for new content: memory_store with pinned: true)
• Unpin: "unpin X" -> memory_unpin

Do NOT ask permission — just store it.

Storage constraints (apply when storing or pinning):

• Pins are short directives, not reference material. Pinned entries are capped at 500 characters. If the content is longer, distill the RULE into <= 500 chars and pin the distillation; keep the full detail as a normal warm memory.
• Store atomic, concise memories at every tier: one fact per entry; split compound observations into separate memories. Long reference content belongs in the knowledge base (kb_ingest), not in memories.

Retrieve (continuous)

On each user message that is a question or task request:

1. Call memory_search with the message, searching warm tier
2. If top score < 0.5, also search cold/knowledge tier
3. Use results to inform your response

Feedback (continuous)

memory_search and kb_search now return a retrievalId alongside results. After you use retrieved memories in your work, call memory_feedback:

• Used it → memory_feedback({ retrievalId, used: true })
• Searched and nothing was relevant → memory_feedback({ retrievalId, used: false })

Report once per search, as soon as you know the outcome. Skip the call only when retrievalId is empty. Do NOT ask permission — just report it. This is what makes the Dashboard "Retrieval quality" metric real.

Pinned directives (continuous)

Pinned directives are re-asserted automatically near the live edge by the per-turn pinned floor (where the host supports it — see the capability matrix in skills/commands/SKILL.md). You do not need to do anything for this.

Additionally: when the user makes a significant task switch (a clearly new piece of work), call session_refresh once. This re-prepends the pinned block and refreshes hot-tier context near the current generation point. On hosts without a per-turn floor (e.g. Cursor) this is the primary way pinned directives stay salient — so do it on task switches there especially.

Session End

1. Call session_context to get current hot tier entries
2. If there are 2+ hot entries, launch the total-recall:compactor agent with the entries as input
3. Parse the agent's JSON decisions and execute them
4. Call session_end for final bookkeeping

Rules

• Let tool calls be visible — users should see that memory is working
• ALWAYS store corrections — highest-value memories
• ALWAYS search warm tier before answering project questions
• NEVER modify host tool files (Claude Code memory/, CLAUDE.md, etc.)