What is an MCP memory server?
An MCP memory server is a Model Context Protocol server whose tools store and recall the agent's own experience. Instead of exposing a SaaS product, it exposes remembering: the agent writes down what it learns and reads it back in a later session, over the same protocol it already speaks.
Any MCP client can connect one. Below: what the tools do, how connecting works, and where the honest limits are.
How is it different from a normal MCP server?
A normal MCP server is a door to a system that exists without the agent: a CRM, a repository, a mailbox. The agent calls it to get something done in that system.
A memory server has no system behind it other than the agent's own past. Nothing in it exists until the agent puts it there, and everything in it is about the agent's work rather than someone else's product. That makes it the one MCP server worth connecting before any of the others, because it is what makes the rest of them worth connecting twice.
Which tools does a memory server expose?
There is no standard set yet, which means the tool names are a design decision rather than a spec. These are the seven Tulimoa Memory exposes, and the reasoning behind each is that an agent should be able to guess what it does from the name alone.
| Tool | What it does | When an agent calls it |
|---|---|---|
| remember | Writes down one durable memory | Right after learning something durable: an ID, a decision, a preference |
| recall | Brings back related knowledge by association, including what was stored today. Without a query: current goal, last checkpoint and open loops | At the start of a task and before finishing. If the memory knows nothing, it says so |
| checkpoint | Records the goal, what is done, next steps and open questions | Once per multi-step task, so a context reset does not lose it |
| forget | Removes a memory or shreds a person from the vault | On request, and for anything that should never have been stored |
| briefing | The summary of what happened overnight | First call after connecting |
| timeline | Reads back a period, including cold storage | When the question is what happened, not what is true |
| link_account | Gets a code a person uses to link the memory in the dashboard. Free | When someone should keep an eye on balance, knowledge and warnings |
How do I connect an MCP memory server?
The same way as any other MCP server. Point your client at the endpoint and pass a bearer token. For Tulimoa Memory that is a POST to https://memory.tulimoa.com/mcp with an Authorization header, and any client that can send one works: Claude Code, Cursor, or your own loop. claude.ai and ChatGPT do not send a custom header; there you sign in with your Tulimoa account instead (OAuth) and pick the memory.
There are two ways to get the key. As a person, you create a memory in the Tulimoa dashboard, start with $1 of credit and copy the key. The second way is the unusual one: your agent gets the key without a human. It calls the register endpoint, receives an HTTP 402 payment challenge, signs a USDC payment on Base, retries, and the key comes back in the response. Then there is no signup form at all, and the wallet is the account.
What does a memory server cost?
Tulimoa Memory charges $0.002 per tool call and that price is fixed. Batch writes cost one call price per 10 units, so importing a long document costs roughly what it costs us to process it rather than a flat nothing. Create it in the dashboard and you get $1 of credit to start, then top up by card from $5.
There is no recurring fee until a memory holds 5,000 things, and from there a daily fee from $0.03 covers the nightly consolidation. If the balance stops covering the fee, only consolidation pauses and calls keep working. An empty balance never deletes anything.
What are the limits worth knowing about?
Three, stated plainly. First, a memory server is only as good as what the agent writes into it; an agent that never calls remember will find nothing to recall. Second, consolidation happens overnight rather than instantly, so knowledge written today is graph knowledge tomorrow, while raw recall works immediately. Third, this is a young category. Tool names are not standardised across providers, so moving between them means remapping, not just changing a URL.
Questions developers ask
Does an MCP memory server work with any MCP client?
Yes. Clients that can send an Authorization header use a key; claude.ai and ChatGPT connect by signing in (OAuth). The protocol is the same one your other MCP servers speak, so there is nothing special to install. Tulimoa Memory speaks Streamable HTTP at https://memory.tulimoa.com/mcp.
Can I run a memory server myself?
You can run one, and several open source projects let you. The parts that are hard to self-host well are the consolidation quality, the privacy handling of personal data and the operational side of a nightly job that must not corrupt anything when it fails halfway.
Can I see what my agent has stored?
Yes, in the Tulimoa dashboard. You see long-term knowledge and single entries, searchable and forgettable one by one, settle contradictions to trusted knowledge, and manage the personal data in the vault. An agent that signed up on its own attaches its memory to your account with link_account.
How is data kept separate between agents?
Each API key belongs to exactly one memory, and every query asserts that boundary. Keys can also carry an agent label, so several agents can share one memory and you can still see which of them learned what.
Where does the data live?
The database is in the EU, in Frankfurt, and the service runs on Cloudflare's network. Consolidation and embeddings run at Mistral, a European provider, so no US model processes the memory. Email addresses, phone numbers, IBANs and known names are replaced by placeholders before anything is stored.
One endpoint, and your agent stops starting over.
Tulimoa Memory is live at memory.tulimoa.com. $0.002 per call, $1 of starting credit in the dashboard, or your agent pays by itself with no signup at all.