REDDIT · REDDIT// 2d agoOPENSOURCE RELEASE

Vektori drops three-layer sentence graph memory layer

Vektori is an open-source memory layer for long-running AI agents that replaces lossy knowledge graph compression with a three-layer sentence graph. By structuring information into Facts (L0), Episodes (L1), and raw Sentences (L2), it enables agents to maintain provenance, track staleness, and resolve contradictions natively within a single Postgres or SQLite database.

// ANALYSIS

Vektori shifts the focus from simple semantic similarity to structured provenance, proving that massive context windows can't replace the need for organized, historical memory.

–Three-layer architecture allows agents to trace abstract facts back to specific source sentences for high-fidelity reasoning.
–Supersession relationships handle conflicting information by archiving old facts rather than overwriting them, preserving a queryable correction history.
–Transitioning from complex Neo4j/Qdrant stacks to a pure PostgreSQL + pgvector implementation drastically simplifies infrastructure for local-first agents.
–Early benchmarks showing 73% on LongMemEval-S suggest a major reliability boost for agents maintaining state over hundreds of sessions.

// TAGS

vektorimemoryagentopen-sourcepostgresvector-dbrag

DISCOVERED

2d ago

2026-04-10

PUBLISHED

2d ago

2026-04-10

RELEVANCE

8/ 10

AUTHOR

Expert-Address-2918