memv-graph

memv-graph

Required reading

1. docs/memv/core-concepts/knowledge-graphs.md — what mem[v] auto-extracts (entities, relationships), how graph built.
2. docs/memv/sdk/graph.md — programmatic query method.
3. docs/memv/core-concepts/search.md — search ALREADY uses graph; may not need explicit graph queries.

SDK surface (verified against docs/memv/sdk/graph.md)

Only one method: client.graph.retrieve_triplets(memory_id, ...) (TS: retrieveTriplets). Returns triplets (entity, relationship, entity) extracted from given memory.

No list_entities, list_relationships, or graph-walk traversal API. Need entities across whole space → query memories via client.memories.search, gather triplets from each result.

When to use this vs plain search

• Plain semantic search (client.memories.search) graph-aware. Traverses entity neighbors automatically. Use for "find memories about X."
• Explicit triplets (client.graph.retrieve_triplets) — use when need:
  • Exact (entity, relationship, entity) structure mem[v] extracted from one memory
  • Audit what mem[v] auto-extracted (debugging or building UI)
  • Build entity-centric panels (per-memory → per-entity rollup in your code)

Need expressible as "find memories about X"? Skip this skill — use memv-search instead.

Hard rules

• Triplets per-memory. Space-wide view → fan out: search memories → call retrieve_triplets on each.
• No writing to graph directly. Entities + relationships derived from memories. Change graph by changing memories.

Skeleton (Python)

# get triplets for a known memory
triplets = client.graph.retrieve_triplets(memory_id="mem_xyz789")
for t in triplets:
    print(t.source, "→", t.relation, "→", t.target)

Don't

• No assuming list_entities / list_relationships / walk methods exist. They don't.
• No modeling own entity/relationship layer on top. Use what mem[v] extracted; if wrong, improve memories.
• No traversing graph in Python for what client.memories.search does server-side.