Open-source Agentic GraphRAG links document knowledge, task memory and decisions in one Neo4j graph

vm324234 · reddit · 2026-10-22

A developer with 3.5 years in RAG/LLM training released the open-source Agentic GraphRAG project (built on Neo4j), addressing a concrete problem: keeping an agent's saved conclusions connected to the evidence behind them.

It links three layers into one "3D graph":

Knowledge — documents, source passages and extracted entities/relations, preserving structure (sections, tables), versions and locations.
Working memory — tasks, progress, messages, tool calls and results via a built-in DeepAgents/LangGraph agent, so sessions can resume.
Decision memory — conclusions with assumptions and open questions, citing exact passages or task events, depending on earlier decisions; reviews can reaffirm or replace decisions without erasing history.

Example: a contract-reviewing agent saves a recommendation with the policy passage it used; when the policy changes, the recommendation and its explicit dependents are flagged for review and excluded from past-decision search until resolved.

Retrieval combines Neo4j full-text + vector rankings with reciprocal rank fusion, with direct passage search as fallback and section/table-aware reads. The built-in agent verifies cited passages were fetched during the run with correct version and permissions (not whether they support the conclusion). Other agents access it via MCP; domain benchmarks are in progress.

Original post →

Open-source Agentic GraphRAG links document knowledge, task memory and decisions in one Neo4j graph

More from coding & agent