Graphiti Documentation — Temporal Knowledge Graph Memory for AI Agents

Graphiti is Zep's open-source Python framework for building temporal knowledge graphs — it calls them Context Graphs — that give agents long-term, queryable memory instead of a growing pile of chat transcripts. The documentation opens with an overview and quick start, then a Configuration section covering the LLM provider (OpenAI by default, plus Anthropic, Google Gemini, Groq and OpenAI-compatible endpoints) and each supported graph store: Neo4j 5.26, FalkorDB 1.1.2, Amazon Neptune with OpenSearch Serverless, and Kuzu, which is documented but deprecated. Core Concepts covers Adding Episodes — data enters as discrete episodes of text or JSON so provenance and chronology survive — plus Custom Entity and Edge Types for domain ontologies, Communities, and Graph Namespacing via group_ids for multi-tenant isolation. Working with Data covers Searching the Graph, CRUD on nodes and edges, and adding fact triples directly, and there is a guide to using Graphiti as the memory layer inside a LangGraph agent. The distinguishing idea is the bi-temporal model: every edge records both when a fact was true and when Graphiti learned it (t_valid and t_invalid), so contradicting information invalidates an old edge rather than overwriting it, and you can ask what the graph believed at a point in time. Retrieval is hybrid — semantic, full-text and graph traversal — with no LLM in the reranking loop, which is how it targets sub-second latency where batch GraphRAG pipelines take seconds. Apache-2.0, Python 3.10+, ~29.6k GitHub stars, with an MCP server and a FastAPI REST service in-tree.

Topic

Agentic

Level

Intermediate

Format

Documentation

Cost

Free

Duration

~3-4 hours for the quick start plus core concepts; ongoing reference after that

Provider

Zep

Hands-on

Yes — code/exercises

Certificate

None

• ✓Model agent memory as a bi-temporal graph of entities, edges and episodes
• ✓Ingest unstructured chat and structured JSON as provenance-preserving episodes
• ✓Define custom entity and edge types to encode a domain-specific ontology
• ✓Run hybrid semantic, full-text and graph search without an LLM reranking loop
• ✓Isolate tenants or individual users with group_id graph namespacing
• ✓Configure Neo4j, FalkorDB or Amazon Neptune as the backing graph store
• ✓Expose your graph to Claude Code or Cursor through the bundled MCP server
• ✓Wire Graphiti into a LangGraph agent as its persistent memory layer

• •The bi-temporal model is the real differentiator: contradicting facts invalidate old edges instead of overwriting them, so history stays queryable
• •The docs include an explicit Graphiti-versus-GraphRAG comparison table — incremental and sub-second against batch and seconds-scale — rather than leaving you to infer it
• •Backed by a published paper (arXiv 2501.13956) reporting 94.8% on Deep Memory Retrieval against MemGPT's 93.4%, and large latency reductions on LongMemEval
• •Ships an MCP server and a FastAPI service in the same repo, so you can try it from an MCP client before writing any integration code
• •States its own caveats up front: Graphiti needs an LLM with reliable structured output, and the sub-200ms figures describe Zep's managed platform, not your own deployment
• •Moving fast and visibly rough at the edges — roughly 271 open issues against 936 commits, so expect to read source for uncommon paths

This page was written from 4 sources, 2 on domains other than help.getzep.com.

1. 1.help.getzep.com — overviewvendor
2. 2.help.getzep.com — llms.txtvendor
3. 3.github.com — graphiti
4. 4.arxiv.org — 2501.13956