RAG that retrieves by reasoning through memory paths

What it does

M-flow is a retrieval engine that organizes knowledge into a four-layer cone graph: Episode, Facet, FacetPoint, and Entity. A query enters through vector search at whichever granularity matches its precision, then the graph propagates evidence along typed, semantically weighted edges to score and return coherent memory bundles. The system is designed to retrieve by reasoning chains rather than flat similarity.

The interesting bit

The project treats the graph itself as the scoring engine, not merely a structural aid. It explicitly distinguishes similarity from relevance: a query anchors at a fine-grained node and the graph routes through facets to the broader episode containing the full context, modeling recall as path-cost optimization rather than vector ranking.

Key highlights

• Unified four-layer cone graph connects episodic summaries, topical facets, atomic facts, and named entities in one traversable structure.
• Retrieval scores by the strongest evidence path, penalizing broad or weak associations while rewarding coherent chains.
• Coreference resolution runs at ingestion, resolving pronouns into concrete antecedents before indexing so cross-turn references remain retrievable.
• Optional face-aware partitioning routes ingestion and queries to person-specific memory datasets via real-time camera recognition.
• Semantic edges carry natural-language descriptions and act as first-class signals in retrieval scoring.

Caveats

• The README references “reported benchmarks” but does not surface any specific numbers or comparative metrics in the visible text.
• Face-aware partitioning depends on an optional external face-recognition repository, so that capability is not self-contained.

Verdict

Worth a look if you are building agent memory or RAG systems where contextual coherence matters more than keyword overlap. Pass if you need a drop-in vector replacement with published benchmark tables.