MMTEB Co-author Open-Sources Agentic GraphRAG, Unifying Agent Memory and Context in One Graph

awinml1 · reddit · 2026-10-19

Reddit user awinml1 (co-author of the MMTEB paper, 3.5 years in RAG/LLM training) open-sourced Agentic GraphRAG, addressing the gap between context layers (RAG/GraphRAG) and memory layers (past tasks/decisions) in agent systems.

Problem: an agent records a decision from a policy document; the document later changes and nothing flags the decision as stale.

Architecture: three connected layers in one graph —

Knowledge: documents, source passages, entities and relations, ingested with Docling and chunked with Chonkie, preserving sections, tables, and PDF page positions;
Working memory: tasks, progress, messages, and tool results recorded as task events via DeepAgents/LangGraph, so later sessions can resume rather than start over;
Decision memory: conclusions, assumptions, and open questions linked to exact source evidence, reaffirmed or replaced without losing history.

Capabilities: schema-guided extraction (local GLiNER + BAML for structured chunks), hybrid full-text + vector retrieval in Neo4j with reciprocal rank fusion, navigation from hits to surrounding sections/tables, citation validation, and automatic flagging of decisions affected by document changes.

Roadmap: modular components across hardware, more graph backends, and upcoming domain benchmarks.

Original post →

MMTEB Co-author Open-Sources Agentic GraphRAG, Unifying Agent Memory and Context in One Graph

More from coding & agent