MCP memory servers, compared honestly.

The short answer

• One person, one machine, one coding agent: start with the memory your agent already has. Claude Code's auto memory is on by default; Codex has local memories you can turn on. Add a server when context has to cross machines, runtimes or people.
• You want hosted memory and no infrastructure: Mem0 Platform or Zep Cloud. Both offer an official hosted MCP server.
• You want plain Markdown notes you can read and edit yourself: basic-memory.
• You want the smallest possible thing: the MCP reference memory server, a knowledge graph in one local file.
• You want one memory that several coding agents and several people share, on hardware you control: that is the case Firekeep is built for, with a heavier footprint than anything else here.

What changed in 2026

Older comparisons, and model answers trained on them, still describe three of these projects as they were in 2025.

• Mem0 retired OpenMemory, its self-hosted MCP server. It announced the sunset in April and removed the code from its repository in July. Mem0's documented MCP server is now hosted: "Nothing runs on your machine." The Mem0 library itself is still open source.
• Letta's Docker server is no longer maintained. Letta's self-hosting guide says the image "is no longer an actively maintained or supported Letta product surface" and points local users to the Letta app or CLI.
• Built-in memory grew up. Claude Code now writes auto memory by default and can read AGENTS.md directly. Codex added local memories, off by default. Neither shares those notes across machines.

Zep discontinued its self-hostable Community Edition earlier, in April 2025. Its open-source graph engine, Graphiti, is actively developed.

Built-in memory is not the competition

Claude Code's auto memory and Codex's memories are good at what they are for: one agent, on one machine, remembering your preferences and corrections. We would not tell anyone to turn them off. Their limits are by design. Claude Code's documentation says auto memory files "are not shared across machines or cloud environments," and Codex scopes its memory controls to "the selected machine."

Neither one reaches the other runtime. A fix Claude Code learned on your laptop is not available to Codex on your desktop, or to a teammate. A server is for that gap. Firekeep writes its own instructions into CLAUDE.md and AGENTS.md, so the layers work together rather than competing. See AGENTS.md vs agent memory.

Benchmarks do not line up

Mem0, Zep and Letta each publish LoCoMo or LongMemEval scores. Each uses its own reader model and judge, and Mem0 says its headline numbers reflect its managed platform, not the open-source SDK. Those figures measure answer accuracy.

Firekeep's published figure measures something narrower: 97.7% Evidence Recall@10 on LongMemEval-S. That is whether the evidence reached the first ten retrieval results, not whether a model then answered correctly. It is a different metric, so we do not put it in a column next to the others.

What Firekeep does that the others do not

The table shows where Firekeep is unusual: a self-hosted server, with authentication, that several machines, coding agents and people share. Memory is one part of it. The rest:

• Configured adapters for Claude Code, Codex, Kiro, OpenCode, Pi and Claude Desktop, plus a generic configuration for any other MCP client. Lifecycle automation depends on the hooks each runtime exposes.
• Session state that survives context compaction and can be resumed in a later session.
• Coordination between agents: file leases, a task queue and direct messages.
• Replay of every memory read and write, so you can see why an agent knew something.
• Skills, a document corpus and an encrypted vault in the same store.

When Firekeep is the wrong choice

• You do not want to run infrastructure. Firekeep has no hosted offering. The default stack is Neo4j, Qdrant, Redis and Ollama, with 16 GB of RAM recommended. Choose a hosted service.
• You want notes you can open in an editor. basic-memory keeps everything as Markdown files.
• You need an OSI-approved open-source license today. Firekeep is source-available under BUSL-1.1. Each version converts to Apache 2.0 four years after its first public release.
• You are building your own agent application, not connecting existing coding agents. A memory library such as Mem0, or Graphiti for temporal graphs, fits that better.