Graphiti
The open-source temporal knowledge-graph engine that powers Zep. Real-time knowledge-graph construction for AI agents — stores entities as nodes, relationships as time-bounded edges, and semantic facts as edge attributes. Repository: [github.com/getzep/graphiti](https://github.com/getzep/graphiti). Apache 2.0 license. Designed to be embedded into agent memory pipelines as the relational substrate underneath retrieval, so it can be adopted independently of Zep's full platform.
Definition
The open-source temporal knowledge-graph engine that powers Zep. Real-time knowledge-graph construction for AI agents — stores entities as nodes, relationships as time-bounded edges, and semantic facts as edge attributes. Repository: [github.com/getzep/graphiti](https://github.com/getzep/graphiti). Apache 2.0 license. Designed to be embedded into agent memory pipelines as the relational substrate underneath retrieval, so it can be adopted independently of Zep's full platform.
Most "memory" libraries store retrieval-time semantic vectors but discard the relationship structure that gives memory its discriminative power. Graphiti is the inverse — it preserves the relational structure as first-class data, then layers semantic-vector retrieval on top of the graph rather than instead of it. Agents can ask "what depends on X" and get a deterministic subgraph rather than a vector-similarity guess.
Embedded temporal knowledge graph for agent memory, real-time graph construction from text, time-aware fact storage with `valid_at` / `invalid_at` semantics, Zep's underlying engine for organizations that want graph memory without the full Zep platform.
Recent developments
- Graphiti MCP Server v1.0 shipped (November 2025). Compatible with Claude Desktop, Cursor, and any MCP-compliant client — turns Graphiti into a drop-in memory backend for the broader MCP ecosystem. The MCP-server form factor is what's driving 2026 adoption growth. Per Skywork — Graphiti MCP Server: Definitive Guide to Agentic Memory.
- Bi-temporal model is the architectural differentiator. Every fact carries two timestamps: valid time (when the fact was true in the world) + transaction time (when Graphiti ingested it). Supports queries like "what was the contract status in March?" — the temporal dimension competitors flatten. Per Codex Blog — Graphiti Temporal Knowledge Graphs for Agent Memory.
- Conflicting facts get invalidated, not deleted. When information changes, old facts are marked invalid; the original record persists with a validity window. Lets agents reason about historical state without destroying prior knowledge — the "what did the user prefer six months ago" query Mem0's ADD-only also targets but solves differently. Per Neo4j Blog — Graphiti: Knowledge Graph Memory for an Agentic World.
- Synthesizes unstructured conversation + structured business data into one graph. 2026 production pattern: Graphiti ingests both chat transcripts and structured operational records, building a unified context graph that preserves the relationship structure missing from flat vector stores. Per GitHub — getzep/graphiti.
- TypeScript port (GraphZep) ratifies the "engine independent of platform" framing. Community-built TypeScript port of the Graphiti paper's architecture confirms the engine is portable beyond the Python+Zep stack — the architecture, not the implementation, is the asset. Per GitHub — aexy-io/graphzep.
- Neo4j published Graphiti as the canonical "graph database for agent memory" pattern. Neo4j's own blog positions Graphiti as the reference architecture for graph-database-backed agent memory — strong validation from the most mature graph-database vendor. Per Neo4j — Graphiti for an Agentic World.
Connections 5
Outbound 4
Inbound 1
depends_on1