jatin salve

ai systems / 2025 - Present

temporal knowledge-graph memory

Graphiti and Neo4j memory for clinical LLM agents, evaluated with a six-stage verification pipeline.

1,682 audited facts6-stage verificationGraphiti + Neo4j

Python · Slurm · Apptainer · Neo4j · Graphiti

Overview

Built a temporal knowledge-graph memory layer for clinical LLM agents using Graphiti and Neo4j. The system emphasizes persistent memory, time-aware retrieval, and evidence that can be traced back to the source record.

Benchmarking

Evaluated memory behavior against 1,682 hand-audited atomic facts using a six-stage verification pipeline.

Verification stage Question
Atomic-fact extraction Is each claim independently testable?
Entity and time alignment Does the fact refer to the correct patient state and period?
Graph retrieval Can the relevant memory be recovered?
Evidence matching Does retrieved evidence support the claim?
Conflict detection Is newer or contradictory evidence present?
Final adjudication Is the memory supported, stale, missing, or incorrect?

Infrastructure and provenance

  • Operationalized experiments with Python, Slurm, and Apptainer.
  • Used Neo4j and Graphiti for persistent graph memory and connectivity.
  • Built deterministic evidence curation, temporal checks, and provenance tracking around hand-audited atomic facts.
  • Contributed to a permission-aware React, TypeScript, and FastAPI application with retrieval, validation, structured logging, Docker, ECS, and CloudWatch.

The result is an evaluation environment where model outputs can be compared against the evidence, configuration, and temporal context that produced them.