Remem vs. Mem0 and Zep: a comparison
Agent memory is becoming a crowded space. Before choosing a memory layer, it helps to understand what each system is actually designed for and where the tradeoffs land. This post is an honest side-by-side. We’ll point out where competitors are stronger.
Data from public documentation as of June 2026.
The deployment question
The first decision is operational: how much infrastructure do you want to run?
Remem’s core is a single Rust binary, remem-server — REST API, storage engine, and MCP server (Streamable HTTP at /mcp) all in one process, started with a single docker compose up. No PostgreSQL, no separate vector store, no graph database. (Prebuilt binaries for a Docker-free install are coming soon.)
Mem0 requires Docker plus an external vector database (Qdrant by default). You’re managing at least two services from day one. Their managed cloud removes this burden but adds cost and a dependency on their infrastructure.
Zep is Go-based and requires PostgreSQL for relational storage and Neo4j for the graph. A full Zep stack is three services: the Zep server, a PostgreSQL instance, and a Neo4j instance. Production deployment is non-trivial.
Qdrant is the simplest competitor operationally — single Docker container, pure vector database. If you only need vector search and no memory lifecycle or graph, Qdrant is excellent.
The graph question
A relationship graph lets your agent follow connections between memories — surfacing context it didn’t know to search for.
Remem builds a graph automatically on every stored memory via auto-discovery. The CSR graph is embedded in-process. find_related traverses it at up to N hops. Included in the free tier.
Mem0 offers graph memory in their Pro plan at $249/month. It uses a separate graph database. The free tier is vector-only with a 10,000 memory limit.
Zep has a temporal graph — it tracks how memories change over time and who said what in which episode. Strong for multi-user applications and session history. Weaker for semantic auto-discovery across arbitrary memories.
Qdrant has no graph at all. Pure vector search.
The lifecycle question
Memory lifecycle is what separates a memory system from a vector database: the ability for memories to expire, decay, be promoted, and be archived.
Remem has the most complete lifecycle: TTL expiration for short-term memories, importance decay over time, automatic promotion to long-term on high access, soft delete (archiving) and hard delete. Background tasks run inside the binary on a configurable schedule.
Mem0 has no lifecycle features in the free tier. The paid tier adds some summarisation.
Zep handles lifecycle at the session level — episodes can be summarised and expired, but individual memory items don’t decay independently.
Qdrant has no lifecycle features (it’s a vector database, not a memory system).
The MCP question
Remem was designed around MCP from the start. Eight tools, three resources, Streamable HTTP in-process (/mcp) plus a stdio companion binary for local clients. Works with Claude Desktop with a one-line config change.
Mem0 has no MCP support.
Zep has no MCP support.
Qdrant has no MCP support.
Where competitors are stronger
Zep is the right choice if you’re building multi-user applications where temporal context (who said what, when, in which session) matters more than semantic graph connections. Their episode model and user/session management APIs are more mature than Remem’s.
Mem0 has a larger ecosystem and more integrations (LangChain, LlamaIndex, Autogen). If you’re deep in the Python AI stack and need drop-in compatibility with existing frameworks, Mem0 will require less custom code.
Qdrant is the right choice if you only need vector search and don’t want any memory layer abstractions. It’s extremely performant and well-documented.
Summary table
| Remem | Mem0 | Zep | Qdrant | |
|---|---|---|---|---|
| Deployment | Single binary | Docker + vector DB | Docker + PG + Neo4j | Docker |
| Language | Rust | Python | Go | Rust |
| MCP-native | ✅ | ✗ | ✗ | ✗ |
| Graph | ✅ Free | $249/mo | ✅ (temporal) | ✗ |
| Memory lifecycle | Full | ✗ | Session-level | ✗ |
| Local dev | ✅ | ✗ (cloud-first) | ✅ | ✅ |
| License | FSL | Apache 2.0 | Apache 2.0 | Apache 2.0 |