Agentic GraphRAG: open-source 3-layer knowledge graph for agent memory and provenance
Aggressive-Judge-448 · reddit · 2026-11-20
Developer released Agentic GraphRAG (ontogr/agentic-graphrag), an open-source framework storing agent-read documents, task history, and decisions in one connected graph so sessions can resume and decisions stay traceable.
Three layers:
- Knowledge layer: documents, passages, entities, relations — passages keep version and location (incl. PDF page positions)
- Working memory: tasks, progress, messages, tool activity; built-in agent uses DeepAgents/LangGraph and resumes across sessions
- Decision memory: conclusions, assumptions, open questions, linked to source passages or earlier decisions; source changes flag dependent decisions for review (e.g., a contract recommendation tied to a changed policy clause)
Tech: Docling ingestion preserves document hierarchy; GLiNER local extraction makes graph building nearly free with no LLM calls; retrieval fuses full-text + vector rankings via reciprocal rank fusion with entity-relationship traversal back to source text. Domain schemas and ontology presets included; built-in agent or MCP integration. Benchmarks coming soon.
More from coding & agent
- Hilton customer service layers identity checks that trip up AI agents — gethackteam · 2026-11-20
- YC-backed Antinuous launches agent platform for running 100x more experiments at 1,000x lower cost — ycombinator · 2026-11-20
- Safe Personal RSI: update gate rejects 20 of 21 risky self-proposed agent updates, cuts attacks to 0% — dair_ai · 2026-11-20
- Skill Store 33 Launches Beta Plugin That Matches Coding Agents to 60,000+ Skills — markjeffrey · 2026-11-20
- Formal verification of x86 code: failed proofs led straight to real glibc bugs — moyix · 2026-11-20
- hermes-browser-bridge: drive your logged-in browser via MCP without a debug port — paq85 · 2026-11-20