Agent memory needs ontology, write paths, and graph-based retrieval
Al_Grigor · x · 2026-07-22
A repost about designing unified memory for agents argues that retrieval is only one layer of the stack. The author says the hard part is the architecture around it, and proposes a four-layer design:
- Ontology: define entities, relationships, merge rules, and how facts change over time.
- Durable write path: separate pipelines for raw document ingestion and transforming documents into structured memory.
- Query / retrieval layer: map natural language to graph queries, expand to connected nodes, find top-K results, and rerank them.
- LLM Wiki / context packing: maintain an index with sources, entities, preferences, and facts, then pack retrieved subgraphs into context.
The diagram ties this to an MCP server with retrieval and ingestion tools such as NL query memory, deep search memory, ingest conversation/file/URL, plus a harness and ingestion pipeline. The overall message is that useful agent memory is a system design problem, not just GraphRAG retrieval.
Related event: Developers Debate the Real Challenges of AI Agent Memory(6 posts)→
More from coding & agent
- Microsoft Foundry Autopilots: agents with their own Entra identity and M365 account — marlene_zw · 2026-10-27
- OpenAI DevDay Berlin: Codex weekly active users up 20x since January — reach_vb · 2026-10-27
- He benchmarks new models by having them build their own copilot — $28 of compute — bucolucas · 2026-10-27
- Founder generates 46 app promo videos in 4 days with Claude, skipping $11,500 in UGC fees — PrajwalTomar_ · 2026-10-27
- Should agents be allowed to change plans mid-task? Reddit debates the drift boundary — wistfulcracker84 · 2026-10-27
- HeyGen + Canva MCP turns slide speaker notes into an avatar presentation video — HeyGen · 2026-10-27