Cognee + Graphiti Memory Evaluation
ADOPT28+ sources across GitHub repos, docs, pricing, comparison articles. Cognee = document pipeline. Graphiti = interaction memory. Together they replace Zep Cloud at zero infrastructure cost.
Key Findings
Cognee (13K stars, Apache 2.0, $7.5M seed): Document-to-knowledge-graph pipeline (Extract, Cognify, Load)
Graphiti (23K stars, Apache 2.0, by Zep): Conversation/interaction-to-knowledge-graph with bi-temporal tracking
KEY: These solve DIFFERENT problems. Cognee = document processing. Graphiti = interaction memory.
Graphiti IS the engine behind Zep Cloud, open-sourced — full replacement for self-hosted
Graphiti supports Kuzu (embedded graph DB, no server, pip install kuzu) — eliminates Neo4j dependency
Kuzu: 'DuckDB of graph databases' — 18x faster ingestion than Neo4j, Cypher query language, works on Windows
FalkorDB Lite also works embedded but Linux/macOS only (no Windows support)
Mem0 graph memory paywalled at $249/mo Pro — basic vector memory only on $19 Starter
Both Cognee and Graphiti have official MCP servers for Claude Code
Recommendation: Replace Zep Cloud with self-hosted Graphiti + Kuzu. Keep Mem0 Starter. Skip Cognee for now.
Action Recommendations
Replace Zep Cloud with self-hosted Graphiti + Kuzu
Keep Mem0 Starter ($19/mo) for vector memory
Skip Cognee for now (document pipeline not needed yet)
Use group_id in Graphiti for per-agent isolation
Monitor Graphiti v1.0 release for API changes
Watch List (from this report)
Graphiti + Kuzu stability — Kuzu backend newer than Neo4j support
FalkorDB Lite Windows support — currently Linux/macOS only
Mem0 Starter graph memory — if graph drops to Starter tier, migration less urgent
Graphiti v1.0 release — currently v0.28.x, major version may change APIs
Zep Cloud free tier changes — if reduced, Graphiti migration becomes urgent