Memory research

Graph retrieval adds the missing hop.

Plain hybrid search is good at finding direct matches. Graph retrieval helps when the answer lives one relationship away: a person, project, decision, or tool connected across several memories.

See all memory features How memory works Issue #559
This is one research track, not the whole Remnic memory system. Remnic also includes extraction, recall injection, hybrid search, entity tracking, lifecycle hygiene, trust zones, procedural memory, local LLM routing, versioning, retention tiers, disclosure controls, project scoping, importers, benchmarks, recall X-ray, and more. These pages explain focused memory tracks inside that broader system.
What changes

The product idea in plain terms.

Start with normal recall

QMD still seeds the result set with BM25, vector search, and reranking. Graph expansion only starts after the strongest direct hits are known.

Walk typed relationships

Entity and relationship metadata become a graph artifact that can connect memories through people, projects, tools, and decisions.

Rerank everything together

Graph-promoted neighbors re-enter the same ranking path as direct hits, so they earn their place instead of bypassing relevance checks.

Already in Remnic
  • Hybrid BM25 + vector recall through QMD.
  • Entity and relationship metadata on memory files.
  • Recall explanation surfaces that show why a memory appeared.
Landed in this track
  • PR #576 added the graph retrieval contract and queryGraph stub.
  • PR #622 added graph edge extraction from memory metadata.
  • PR #629 implemented Personalized PageRank, PR #640 wired the flag-gated tier, and PR #644 added the retrieval-graph A/B benchmark.
References
Related Remnic pages