memento-mcp: the agent memory you can actually see and edit

Short version: Most AI memory is a black box — you trust the agent stored the right thing, but you can't easily see or fix what it remembers. memento-mcp is a knowledge-graph memory with first-class inspection tools: entities, relations, and observations you can list, audit, and edit by hand. Install via npm, 10 minutes. By gannonh, MIT licensed.

What is memento-mcp?

memento-mcp is an MCP memory server built around a knowledge graph with three primitives: entities (people, projects, companies), relations (employs, competes-with, invested-in), and observations (raw facts attached to entities). Unlike simpler memory tools, it ships with inspection tools as a headline feature — your agent can list every entity it has stored, show how they're connected, and let you edit wrongs.

The philosophy: memory you can't audit isn't memory you can trust. Most founders won't put an agent in front of real workflows until they can verify what it thinks it knows. memento-mcp is built for that use case.

Who this is for

Power users who want to inspect and correct what their agent believes.
Teams doing real work with AI agents where a wrong belief has a real cost (wrong customer, wrong deal term, wrong decision).
CEOs who want a "what does my agent know about me?" view they can actually read.
Anyone who has been burned by an agent confidently misremembering something.

Skip this if

You never want to look inside the memory — trust-the-black-box users should use memory-mcp (simpler) or mem0 (production-scale). memento-mcp pays for inspection with a bit more setup.

What problem it solves

Every AI founder I know has had this moment: the agent says something confident and wrong. "The customer's contract expires in June." No it doesn't — it's July. Where did the agent get that? Can I fix it? How do I stop it from happening again?

memento-mcp answers those questions directly. Every entity, relation, and observation is inspectable. Wrong beliefs get caught, fixed, and committed — the same way you'd fix a typo in a wiki. Over time, your memory accrues trust because you've seen inside it.

How to install it (plain English)

Install globally. npm install -g memento-mcp.
Add to your MCP config. { "mcpServers": { "memento": { "command": "memento-mcp" } } }.
Restart your client.
Teach it some entities. Ask your agent: "Remember that my company is Linea, founded 2025, 30 people, DTC skincare. My co-founder is Amara." The agent uses memento's tools to create entities and relations.
Inspect the graph. Ask: "Show me every entity you've created, with their relations." memento-mcp has a dedicated inspection tool — you'll get a clean printout you can review.

Full walkthrough: /memory/tools/memento-mcp.

What you can do with it (for a non-technical founder)

Audit what your agent knows — "show me everything you believe about my company" produces a readable report.
Fix wrongs before they act — spot a mis-categorized customer; fix it with a "rename/merge/delete" command.
Build a living org chart — entities + relations = a graph of your team, customers, investors, vendors.
Export to git — snapshot the graph as JSON and commit it. You now have a versioned record of what your agent knows.
Trust, then delegate — once the memory is clean, you can let the agent act on it with higher confidence.

What CLO adds on top

memento-mcp gives you a graph of what your agent believes. Cognition CLO gives you a model of what your humans are retaining. You'll often want both. The agent-memory layer answers "is my assistant correct?" The CLO layer answers "is my team retaining the right things?" Different problems, both worth solving.

FAQ

How is this different from mem0?

mem0 is LLM-managed — it decides what to extract and store. memento-mcp is structured-first — entities and relations are explicit, and you can inspect them directly. Use memento-mcp when you care about auditability.

Can my agent use both memento and mem0?

Technically yes (both register as independent MCP servers). Pragmatically, pick one — overlapping memory layers cause confusion.

What counts as an "entity"?

Anything you'd track: people, companies, projects, customers, meetings. Think of it as the noun list of your business.

Does it scale to thousands of entities?

For a typical founder's use case (a few hundred entities — people, customers, projects), performance is fine. For millions, look at graph databases like graphiti (Neo4j-backed).

Can I import from another memory tool?

memento-mcp supports JSON import/export. A one-off script can migrate from mem0 or memory-mcp; the repo has examples.

What if my agent stores something wrong?

Ask it to fix via memento's edit tools: "Update the entity 'Acme Corp' — correct founding year to 2023." The agent calls the tool; the graph updates; you can re-inspect to confirm.

Ready to install? Full walkthrough at /memory/tools/memento-mcp. Credit to @gannonh for building and maintaining the project — star the repo if the audit-first model earns your trust.