Best MCP Memory Servers in 2026: 7 Options Compared

The best MCP memory server for most people is Hjarni, a knowledge base with a built-in MCP server that Claude and ChatGPT can search, read, and write. Developers building their own agents should look at Mem0 or Supermemory instead. Local-first users should look at Basic Memory. Here are all seven compared.

Every option below speaks MCP, the open protocol Claude, ChatGPT, Cursor, and most AI tools now use to reach outside data. They solve the same problem: your AI forgets everything between sessions. They solve it in very different ways.

The short version

Server	Best for	Memory model	Price
Hjarni	People who want a knowledge base they can read	Markdown notes, folders, tags	Free to start, Pro $10/mo
Mem0 / OpenMemory	Developers building agents	Extracted memory fragments	Free tier, paid from $19/mo
Supermemory	Developers who care about latency	Extracted memories via API	Free tier, paid from $19/mo
Zep	Enterprise agent teams	Temporal knowledge graph	Open source engine, hosted cloud
Letta	Builders who want a full agent runtime	Tiered agent memory	Open source, hosted cloud
Basic Memory	Local-first tinkerers	Local Markdown files	Free local; cloud $15/seat
Knowledge graph server	Trying MCP memory in five minutes	Local JSON graph	Free, open source

Pricing is as listed by each vendor in June 2026. Check their sites before deciding, these numbers move.

How I picked

One rule: would I trust it as the long-term home for context I never want to re-explain? That favors memory you can read, edit, and export. It penalizes black boxes, even clever ones. I build Hjarni, so it is named first. The caveats on every entry, including ours, are real.

1. Hjarni

Hjarni is a note app with a built-in MCP server. You write Markdown notes in folders with tags. Claude and ChatGPT connect with one URL and can search, read, and create notes. Your AI's memory is just your notes, so you can open them, fix them, and reuse them yourself.

Where it wins. The memory is human-readable. Nothing is extracted or summarized behind your back. The same notes serve every client: Claude, ChatGPT, Cursor, Copilot. You can set per-folder LLM instructions that tell your AI how to write, tag, and file new notes. The free tier is real and includes full MCP access. No credit card required.

Where it falls short. Nothing is automatic. You or your AI write notes deliberately; Hjarni does not mine your conversations for facts. And it is not an embeddable memory API. If you are building memory into your own product, Mem0 or Supermemory fit better.

Price. Free to start. Pro is $10 per month for unlimited notes and file attachments.

2. Mem0 / OpenMemory

Mem0 is a memory layer for AI applications. It watches conversations, extracts facts, and serves them back through an API or MCP. OpenMemory is its local-first MCP server that shares memory across Claude Desktop, Cursor, and other tools on your machine.

Where it wins. Automatic extraction at scale. If you are building an agent product and want memory without designing it yourself, Mem0 is a common default choice. It ships SDKs, graph memory on higher tiers, and editor plugins for Claude Code and Cursor.

Where it falls short. The memories are fragments optimized for retrieval by software, not documents you would read. The free Hobby tier caps retrieval calls. The paid ladder climbs fast: Starter at $19, Growth at $79, Pro at $249 per month.

Price. Free Hobby tier (10K memories, capped retrievals). Paid plans at $19, $79, and $249 per month, as listed in June 2026.

3. Supermemory

Supermemory is a memory API. It ingests content, extracts and updates facts, handles contradictions, and serves recall fast. The team leans hard on speed. Its pricing page claims sub-300ms p50 latency for search.

Where it wins. Latency and room to prototype. The free tier includes $5 of monthly usage and the $19 Pro plan includes about $20. The MCP server and Claude Code plugin make it easy to wire into coding workflows.

Where it falls short. Like Mem0, it is developer infrastructure first. Pricing is usage based, so costs take some math to predict. Extraction is automatic, which means you trust the system about what was worth remembering.

Price. Free tier with $5 of monthly usage included. Pro $19, Max $100, Scale $399 per month, as listed in June 2026.

4. Zep

Zep builds a temporal knowledge graph from your data using its open source Graphiti engine. Facts get validity windows, so the system knows not just what is true but when it changed. It posts strong results on long-term memory benchmarks.

Where it wins. Changing state. If your agent tracks users whose situation evolves, the temporal graph is genuinely ahead of flat memory stores. Enterprise features and an open source core.

Where it falls short. It is built for engineering teams shipping agent products. As a personal memory for Claude or ChatGPT it is heavy machinery for a small job.

Price. Graphiti is open source. The hosted cloud is usage based; check their pricing page.

5. Letta

Letta grew out of the MemGPT research project. It is not a bolt-on memory store but a full agent runtime where memory management is part of the operating model, with core, archival, and recall tiers.

Where it wins. Depth. If you are building agents from scratch and want memory designed in rather than added on, Letta's architecture is the most considered of this list.

Where it falls short. You adopt a platform, not a plugin. For adding memory to an existing setup, or to your daily Claude use, it is the longest road here.

Price. Open source, with a hosted cloud offering.

6. Basic Memory

Basic Memory stores your AI's memory as plain Markdown files on your own machine, readable and editable in Obsidian or any editor. The local version is free and open source. It connects to Claude Desktop, Codex, and Cursor over local MCP, and a hosted cloud with team plans now exists too.

Where it wins. Total ownership. Files on disk, no server, no account. The philosophy is close to Hjarni's: memory should be notes a human can read. If you want that fully local, this is the pick.

Where it falls short. Running it locally cuts both ways. Remote clients like ChatGPT on the web cannot reach your machine, and sync across devices is your problem. The hosted cloud solves that, but at $15 per seat per month it costs more than most personal options here.

Price. Local version free and open source. Cloud from $15 per seat per month.

7. The official knowledge graph server

The reference memory server from the Model Context Protocol project stores entities and relations in a local JSON knowledge graph. It is the simplest possible way to see MCP memory work.

Where it wins. Five-minute setup, zero cost, and it teaches you what MCP memory actually does. Good default for experiments.

Where it falls short. One flat file on one machine. No real search, no UI, no sync. It is a demo that people press into production, and it shows at a few hundred entities.

Price. Free, open source.

The best MCP memory server, by use case

Pick by what you are actually doing.

• You want Claude and ChatGPT to remember your work. Use Hjarni. Write notes once, connect with MCP, and your AI reads them in every conversation. That is what long-term memory for AI looks like in practice.
• You are building an AI product and need memory infrastructure. Mem0 or Supermemory. We compared them directly in Mem0 vs Supermemory vs Hjarni.
• Your agents track state that changes over time, at company scale. Zep.
• You are designing agents from first principles. Letta.
• You want everything local and on disk. Basic Memory.
• You just want to see MCP memory work today. The reference knowledge graph server.

And if you are still weighing approaches beyond MCP servers, from built-in ChatGPT memory to CLAUDE.md files, start with the best ways to give your AI memory.