10 Consensuses on Agent Memory Architecture

pauliusztin · reddit · 2026-07-15

The author systematically reviewed the implementations of Cognee, Graphiti, and Neo4j agent-memory, concluding that despite different product forms, their core architectures are highly similar.

Shared Architectural Principles

Coexistence of Knowledge Graphs and Vector/Text Retrieval: Rather than choosing one, they are fused within the same storage layer.
Ontology/Data Model is Core: All extraction and querying revolve around a unified schema. The author mentions Neo4j's POLE+O (Person, Object, Location, Event, Organization) along with preference and fact nodes.
Single Database Preference: Placing text search, graphs, and indexes in one database reduces operational complexity; MongoDB is suggested as a viable single-DB solution.
LLM-driven Extraction: Nodes and edges are extracted from chunked inputs, using batching and cheaper models to control costs.
Separating Resolution and Deduplication: Resolve names first, then merge identities, because "false merges" are irreversible.
Nightly Cleaning/Re-deduplication: Scheduled reruns of dedup act like memory "sleep consolidation."
Logs vs. Single Collections: Append-only offers versioning and temporality but costs more memory; single collections are more practical when versioning isn't needed.
Three Query Methods: Graph search, agentic search, and on-the-fly generated LLM wikis.
Closed-source APIs First, Open-source SLMs Later: Switch when cost, latency, or privacy demands it.
Exposure via MCP and Skills: Package it as a memory app rather than a database wrapper, providing only essential ingest/write primitives.

The author concludes by noting that their biggest challenge in building agent memory is balancing depth and breadth in ontology design.

Original post →

10 Consensuses on Agent Memory Architecture

More from coding & agent