Skip to content

Graphiti / Zep

Graphiti is an open-source framework for building temporal context graphs for AI agents. Unlike static knowledge graphs or RAG pipelines, Graphiti tracks how facts change over time: each entity and relationship carries a validity window, old facts are invalidated rather than deleted, and every derived fact traces back to the episode that produced it. Retrieval combines semantic, keyword (BM25), and graph traversal in a single hybrid query.

Zep is the managed, enterprise-grade platform built on top of Graphiti. It provides governed, production-ready context graph infrastructure pre-configured retrieval, sub-200ms latency targets, user/thread management, a developer dashboard, and SDKs for Python, TypeScript, and Go.

Sources: getzep/graphiti · getzep.com

Ikidna’s context requirements centre on structural codebase knowledge that is accurate, current, and accessible to a distributed swarm. Graphiti addresses the most demanding parts of that problem:

  • Temporal correctness facts carry validity windows and are automatically invalidated when superseded, which directly supports Ikidna’s goal of keeping agents in sync with a changing codebase without full re-indexing
  • Incremental ingestion new episodes integrate immediately; only changed surfaces are reprocessed, satisfying Ikidna’s scale requirement
  • Hybrid retrieval a single query spans semantic similarity, keyword matching, and graph traversal, so agents do not need to chain multiple retrieval calls
  • Provenance every fact links back to the source episode, which maps naturally onto Ikidna’s review and audit requirements
  • Pluggable graph backends supports Neo4j, FalkorDB, Kuzu, and Amazon Neptune, aligning with the pluggable-storage requirement in the context engine spec
  • MCP server included Graphiti ships an MCP server, giving harness-level integration a documented path without custom tooling
  • Apache 2.0 licence no commercial use restriction

The Zep managed offering adds the governance, SLA, and observability layer that a production Ikidna deployment would require if self-hosting the graph engine is not preferred.