Mem0 vs Zep vs OctaMem: choosing a memory layer

Mem0, Zep, and OctaMem all give AI agents memory that persists across sessions, but they take different approaches. Mem0 automatically extracts and stores memories with an LLM pipeline and has a large open-source community. Zep is built on a temporal knowledge graph and specializes in reasoning about how facts change over time. OctaMem stores memory as explicit, typed records (semantic, episodic, procedural) with provenance, logs every access, and exposes it through a production remote MCP server. The right choice depends on whether you value automatic extraction, temporal reasoning, or typed auditability.

Side-by-side comparison

The core difference: how memory is written

The clearest way to tell these three apart is how a memory gets created — from fully automatic LLM extraction to explicit, user-declared typed records:

Mem0: automatic extraction, big community

Mem0(“mem-zero”) is a universal memory layer that sits between your app and the LLM and automaticallyextracts facts: a two-phase pipeline where an LLM proposes candidate memories and then decides whether to add, update, delete, or ignore them. You don’t author records — the system infers them. It’s open source under Apache 2.0 with one of the largest communities in the space, REST/Python/JavaScript SDKs, and a platform MCP server. Graph memory (“Entity Linking”) is gated to its Pro tier.

Choose Mem0 if

You want memory to “just work” with minimal wiring, you value a large open-source ecosystem, and automatic extraction plus token-efficiency matter more than controlling exactly what gets stored.

Zep: temporal knowledge graph

Zep is built around Graphiti, an open-source temporal knowledge graph. Its distinguishing feature is a bi-temporal model: every fact carries timestamps for when it was recorded and when it was true in the real world, and new contradicting facts invalidateold ones instead of deleting them — preserving history with provenance. That makes Zep strong at temporal reasoning (“what did we know on this date?”). Note its MCP server is officially experimental, and the full self-hostable app was deprecated in 2025 (you can still self-host Graphiti with a graph database).

Choose Zep if

Your use case hinges on how facts change over time, you want a knowledge-graph model, and temporal reasoning is a first-class requirement.

OctaMem: typed, auditable, remote MCP

OctaMem stores memory as explicit typed records — semantic, episodic, and procedural — each with provenance, and logs every read and write by default. Rather than inferring what to keep, you (or your agent) declare memories through a clean API or the hosted remote MCP server, which works from Claude Desktop, Claude.ai, and Cursor with no local install. It’s a managed service with cloud, VPC, and on-premise deployment, aimed at teams that need memory to be typed, traceable, and removable.

Choose OctaMem if

You need auditability and provenance as defaults (not an enterprise add-on), you want typed memory you control, and a production remote MCP server plus VPC/on-premise deployment matter — common in regulated industries like healthcare, finance, and legal.

How to choose

There’s no single winner — the fit depends on your priority:

Whichever you pick, the underlying concepts are the same — see the four types of AI agent memory for the model all three build on, or add persistent memory to Cursor and Claude to try it hands-on.

Frequently asked questions