name	backant-memory
description	Use when starting work on any task with possible history (recall first), after completing or failing an attempt (write episode), after verifying durable knowledge (write LTM), or when a repeatable runbook emerges (propose procedure). Covers all mcp__backant-memory__* tools.

backant-memory — persistent vectorized memory

Repo-scoped, local-first memory from the backant-memory MCP server (per-session, scoped to this repo). A background service keeps the local embedding runtime warm. Tool prefix: mcp__backant-memory__.

Session start

procedure_sweep once per session (marks stale runbooks).
task_state_read (no args) to list active epics; with epic_id to resume one.

Before deciding or building

memory_recall {cue, k?} — hybrid BM25+vector+recency retrieval. ALWAYS do this before re-deriving anything that may have history.
memory_recall_with_edges {cue, edge_depth?} — when consequences/relations matter.
memory_pattern_check {domains} — check for repeated failure patterns before retrying an approach.
procedure_grounding {trigger, action_type} — fetch runbooks for the action; follow "follow-this" results, verify "draft-verify" ones.

After acting

memory_write_episode — every meaningful attempt: situation, action, expected vs outcome, evidence.
memory_write_stm {type, content, sources} — observations/hypotheses/anomalies worth keeping.
memory_write_ltm {type, content, sources, reason} — ONLY verified durable knowledge.
procedure_outcome {procedure_id, outcome, ...} — whenever a runbook was applied.
procedure_propose — after a VERIFIED success worth repeating; list depends_on_paths honestly.

Curation (when reviewing memory, not every session)

memory_reinforce / memory_bump_verdict_boost — memory proved useful.
memory_revise_ltm — content is outdated; state the reason.
memory_promote / memory_demote — STM↔LTM transitions with reasons.
memory_edge_propose → review → memory_edge_approve / memory_edge_reject; list open ones with edges_pending.
memory_decay_sweep — periodic hygiene only.

Long-running work

task_state_write — rewrite the durable plan for an epic at every significant state change.

backant-memory — persistent vectorized memory

Session start

procedure_sweep once per session (marks stale runbooks).

task_state_read (no args) to list active epics; with epic_id to resume one.

Before deciding or building

memory_recall {cue, k?} — hybrid BM25+vector+recency retrieval. ALWAYS do this before re-deriving anything that may have history.

memory_recall_with_edges {cue, edge_depth?} — when consequences/relations matter.

memory_pattern_check {domains} — check for repeated failure patterns before retrying an approach.

procedure_grounding {trigger, action_type} — fetch runbooks for the action; follow "follow-this" results, verify "draft-verify" ones.

After acting

memory_write_episode — every meaningful attempt: situation, action, expected vs outcome, evidence.

memory_write_stm {type, content, sources} — observations/hypotheses/anomalies worth keeping.

memory_write_ltm {type, content, sources, reason} — ONLY verified durable knowledge.

procedure_outcome {procedure_id, outcome, ...} — whenever a runbook was applied.

procedure_propose — after a VERIFIED success worth repeating; list depends_on_paths honestly.

Curation (when reviewing memory, not every session)

memory_reinforce / memory_bump_verdict_boost — memory proved useful.

memory_revise_ltm — content is outdated; state the reason.

memory_promote / memory_demote — STM↔LTM transitions with reasons.

memory_edge_propose → review → memory_edge_approve / memory_edge_reject; list open ones with edges_pending.

memory_decay_sweep — periodic hygiene only.

Long-running work

task_state_write — rewrite the durable plan for an epic at every significant state change.