← Back to all projects

Graphiti

Agent memory as a knowledge graph that records when each fact was true, not only that it was

OfficialApache-2.0
Stars
32.4k
Forks
3.3k
Open issues
529
Last commit
28 Oct 2026

What is Graphiti?

Most memory layers overwrite: a new fact replaces the old one and last year's answer becomes unrecoverable. Graphiti instead marks the old relationship as no longer holding, with the dates attached, so both what is true now and what was true then remain queryable — and it separates when something happened from when the system learned about it, which is what keeps late-arriving information from rewriting history. Retrieval combines vector similarity, full-text search and graph traversal. The costs are concrete: a graph database has to run alongside your stack, and every ingest calls a model to extract entities and relationships.

What can you do with Graphiti?

  • Record when a fact stopped being true — Edges carry temporal metadata and are invalidated rather than overwritten, so a question like which supplier was in use last quarter still has an answer after the supplier changed.
  • Separate when it happened from when you heard it — Event time and ingestion time are modelled independently, which means information that arrives out of order is placed correctly instead of appearing to be the newest state of the world.
  • Ingest in episodes, not as a stream — Data enters as discrete episodes that keep their provenance, so entities and relationships are extracted incrementally and any fact can be traced back to the episode it came from.
  • Retrieve three ways at once — Vector similarity, full-text search and graph traversal are combined without a model reranking pass, and results can be ordered by distance from a chosen node rather than by score alone.
  • Name the entity types your domain has — Custom entity types steer extraction toward the categories that matter in your data instead of leaving it to generic people, places and organisations.

Before you choose Graphiti

  • It needs a graph database — Neo4j, FalkorDB or Amazon Neptune — running beside whatever you already operate, which is a second stateful system to size, back up and upgrade.
  • Building the graph means a model extracts entities and relationships from every episode, so writes carry a per-ingest model cost that a plain vector store does not.

Star history

21 Aug to 30 Oct · +2.2k

30.2k32.4k

Frequently asked questions

Is Graphiti free for commercial use?

Graphiti is released under the Apache-2.0 licence — OSI-approved open source, which permits commercial use.

How can Graphiti be deployed?

Graphiti is available as Self-hosted / Runs locally.