Zep
Long-term memory for AI assistants and agents — temporal knowledge graphs over conversation history.
About
Zep is a memory layer for assistants and agents that need to remember conversations across sessions. It ingests dialogue, builds temporally-ordered knowledge graphs with its Graphiti engine — facts get versioned and contradictions resolved when the user changes their mind — and serves it back through a REST API with Python and TypeScript SDKs. Benchmarks published against GPT-5-class judges put it at 90%+ on LongMemEval and 94.7% on LoCoMo, with retrieval latency around 100ms.
There is a managed Zep Cloud tier and a self-hostable option, plus SOC 2 Type II and HIPAA attestations that make it the compliance-friendly pick among memory tools. The honest trade-offs from independent comparisons: the graph smarts arrive via Graphiti as a dependency rather than a unified architecture, MCP support is limited, and pricing scales with usage — it is chat-history memory first, not a general knowledge pipeline.
What it can do
- Long-term conversation memory with temporal awareness
- Knowledge graphs via Graphiti: fact versioning, contradiction resolution
- REST API plus Python and TypeScript SDKs
- Multi-tenancy and session isolation for team deployments
- SOC 2 Type II and HIPAA attestations; BYOC option
- ~100ms retrieval latency; auto-search mode cuts token usage
Setup
- Sign up for Zep Cloud or self-host the open-source server via Docker.
- Install an SDK:
pip install zep-cloud(or the TypeScript SDK). - Create a user and session, then add conversation turns via the API.
- Query memory with graph search or the lower-token auto-search mode before each agent turn.
- Tune fact extraction and contradiction handling in the dashboard as your schema settles.
Prerequisites
A Zep Cloud account or a self-hosted Docker deployment; an LLM provider key for the agent itself.
Use cases
Works with
Trust score — 8/10
Scored 8/10 (metrics tier, metrics-v1). Evidence: 4,960 GitHub stars (GitHub API), Apache-2.0, pushed Oct 14 2026; two independent 2026 comparisons corroborate the Graphiti-based architecture, the benchmark figures, and the SOC 2/HIPAA positioning. Strengths: purpose-built for long conversation memory, strong published benchmarks, the compliance story is real. Docked for: (1) the retrieval engine around Graphiti is closed in the managed/BYOC path — auditors can't read all of it; (2) chat-history-centric, not a general ingestion pipeline; (3) usage-based pricing adds up on large teams. License: Apache-2.0 (repo); managed service terms apply to Cloud.
★ 4,960 GitHub stars
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 10.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Adoption 7: starScore from 4,960 stars on getzep/zep
- · Maintenance 10: pushed 2026-10-14 (recency 10)
- · Openness 9: Apache-2.0 repo (0.7*9, docked for closed retrieval engine in managed/BYOC path) + docs verified (0.3*10)
- · Manual 8: two independent 2026 comparisons corroborate the Graphiti architecture, benchmark figures (LongMemEval 90.2%, LoCoMo 94.7%), and SOC 2/HIPAA positioning; docked for closed engine in BYOC, chat-history-centric scope, usage-based pricing
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.