wrongsynapse-memory

name	wrongsynapse-memory
description	Persistent project memory for coding agents — query WrongSynapse before design decisions or file edits, record observations during work, consolidate candidates when a task completes. Use when working in any repo indexed by WrongSynapse or when the wrongsynapse CLI/MCP tools are available.

WrongSynapse Memory Skill

You have a local, persistent memory layer for this project: WrongSynapse (npm: wrongsynapse). It gives you tri-hybrid retrieval (semantic + FTS5/BM25

• knowledge-graph) over the codebase structure, decisions, and episodic notes — 100% local, zero cloud. This skill wires the memory reflex into your workflow so knowledge survives session boundaries.

Setup (one-time, if tools are missing)

• MCP client config (Claude Code / Claude Desktop / Cursor / any MCP host):

{
  "mcpServers": {
    "wrongsynapse": {
      "command": "wrongsynapse",
      "env": { "SYNAPSE_DB_PATH": "/absolute/path/to/synapse.db" }
    }
  }
}

• If the synapse_* tools are not visible, run wrongsynapse --index . once in the workspace to populate the graph, then reconnect the MCP client.

When to use

• Before non-trivial work: query memory for prior decisions, gotchas, and related symbols touching the files you are about to edit.
• During work: record surprises, root causes, conventions you had to discover — cheap writes, no graph pollution.
• At task end: consolidate — promote keepers, discard noise.

Workflow

1. Recall — before designing or editing: synapse_hybrid_query { query: "<feature/area>", scopes: ["proj:<name>"], limit: 5 } Read ranks (fts/vector/graph) to see why each hit matched. Trust stale entries only after checking the anchored file still says so.
2. Orient — find the real structure: synapse_graph_neighbors { entity_id: "proj:<name>/file:src/auth.ts", depth: 1 } Callers, callees, anchored memories, and hierarchy in one hop.
3. Work — do the coding task.
4. Observe — for each durable insight (root cause, convention, decision, trap): synapse_record_observation { content: "[domain-term: <term>] <one self-contained sentence>", scope_path: "proj:<name>/file:<path>", confidence: 0.85 } Tag domain terms ([domain-term: pushPrefixClause]) so glossary queries find them; one coherent fact per observation.
5. Consolidate — at task end review the pool: synapse_list_candidates { status: "pending" } → synapse_promote_candidate { candidate_id, target_scope } for keepers; synapse_discard_candidate { candidate_id } for noise. Discard is terminal — a discarded candidate can never be promoted.

Scope URIs (addressing scheme)

• proj:<name>/pkg:<pkg>/dir:<path>/file:<path>/sym:<symbol> — hierarchy; proj:<name>/commit:<sha> for git history.
• Prefix filters are boundary-aware: proj:app matches proj:app/... but NEVER the sibling proj:app2/....

Rules

• Never guess a symbol's neighborhood — one graph_neighbors call beats re-deriving call graphs by hand.
• Record the why, not the what; the graph already knows the what.
• Zero-cloud: the embedding model is local; if it is missing, semantic search degrades gracefully to lexical + graph (still useful — keep querying).
• Don't spam observations: if a fact is obvious from the code, the graph already has it.