Connecting your company memory to Claude over MCP
Claude remembers you. It keeps your preferences, your style, your ongoing projects. What it does not remember is your company: the pricing decision three people made in February, the reason you ruled out that supplier, what the last campaign taught the team. That knowledge belongs to no single conversation and no single person, and that is precisely why no assistant's personal memory carries it.
The clean answer to this problem fits in one protocol: MCP. The company's memory lives in a dedicated system, outside the assistant, and Claude queries it during the conversation, the way it would consult a colleague who knows. This guide shows how to connect that in five steps, what it changes in practice, and what to demand from any MCP server before trusting it with a company's memory.
Why MCP rather than Claude's own memory
Three reasons, in order of importance.
Company memory is shared, assistant memory is personal. What Claude retains from your conversations serves only you. A company decision must be readable by your co-founder, by the person who joins next month, and by whatever AI tool they use, which may not be Claude.
The memory outlives the subscription. What an assistant learns about your business stays with it. A memory connected over MCP lives in a system you control: switch assistants, it stays. The same server connects to ChatGPT, Cursor or Gemini, which read and enrich the same memory without any of them owning it.
Answers come with sources. A company memory worth the name ties every claim to its evidence: the decision, its date, what motivated it. Claude then answers by citing, instead of reconstructing from thin air a context it never had.
The connection in five steps
The steps below are Verbasil's, but their shape holds for any serious memory MCP server: a token generated on the memory side, a connector declared on the client side.
- Open your memory tool, Integrations section, AI tools.
- Generate an access token, choosing its scope: read-only, or read and write. Copy it immediately, it is shown only once.
- In Claude, open the settings, then the connectors.
- Add a custom connector with the server's MCP address and your token.
- Open a conversation and ask a control question: "What does my memory contain?" If the tools appear and the answer cites real items, you are connected.
Two details that save support tickets: Claude Desktop and Claude Code use the same connector, one token covers both. And if you connect several AI clients, give each its own token rather than sharing one: call limits are set per token, and two clients on the same token steal each other's rate.
What you ask Claude once connected
The connection is only worth the questions it makes possible. The ones that come up most:
- Before deciding: "What have we already decided on pricing, when, and why?" The answer cites the dated decision and its reasons, instead of reopening the debate from zero.
- Before a meeting: "Give me the context on this contact before I write to them." History, commitments made, what is still open.
- Before repeating a mistake: "What did we learn the last time we tried this?"
- To pick the thread back up: "What is still open this week, and what is waiting on me?"
The common pattern: each of these questions would otherwise have cost a colleague interruption, a dig through email, or a decision retaken blind.
Writing back: the rule that makes it sustainable
An MCP server with write access lets Claude offer to record what was just said: a decision settled in conversation, a learning, a meeting summary. That is the virtuous direction, the memory feeds itself during the work instead of demanding a data-entry discipline.
But the rule that separates a healthy system from one that degrades: an AI write is never applied directly. At Verbasil, it lands in Sources and waits for a human to review it. An assistant writing into a company's memory with no net would make it unusable within a month: a memory is worth what its validation is worth.
What to demand from any memory MCP server
Whether you choose our tool or another, these five requirements are not negotiable.
- Scoped tokens. Read-only or read and write, your choice, and tied to one specific memory, not to the whole account.
- Immediate revocation. A revoked token must stop answering instantly, and what is in the memory must stay there.
- Per-token call limits. In our case, 30 calls per minute and 400 per hour: an agent left looping must not be able to consume your memory without end.
- The server sees only the questions. The MCP server receives what the assistant asks it, never the rest of your conversation. If a vendor cannot state that in writing, that is an answer in itself.
- Human validation on every write. See above. Without it, a write token is a risk, not a feature.
These requirements are the start of a wider subject, MCP server security for companies, which we treat separately.
The limits, said plainly
Claude is the simplest client to connect, others vary. ChatGPT goes through its developer mode, which adds several steps; the full walkthrough is on the Claude and ChatGPT pages.
The memory returns only what it holds. An MCP server plugged into an empty memory answers with emptiness. The connection comes after the material: decisions, meetings, notes. The guide how to write your notes covers that side.
The reading direction is one-way. Claude queries the memory, the memory does not crawl your Claude conversations. What should enter the memory enters through your real sources or through a validated proposal, never by silent ingestion.
Book a memory audit: free, 30 minutes, we look together at what your memory should contain before you even connect it.
FAQ
How do you connect a company memory to Claude?
Over MCP, in five steps: generate an access token in your memory tool choosing its scope, open Claude's settings then connectors, add a custom connector with the server's MCP address and the token, restart the client, and ask a control question. Claude Desktop and Claude Code share the same connector.
What is a company memory MCP server?
A server that exposes the company's memory, its dated decisions, learnings and contacts, to any MCP-compatible AI assistant, as read and write tools. The memory lives outside the assistant: several AIs read and enrich the same one, and none owns it. That is the difference with an assistant's personal memory, which retains only you and stays with the vendor.
Can Claude write into the company memory?
Yes, if the token has write scope, and on one condition that should be non-negotiable: no write applied without human validation. Claude offers to record a decision or a learning, the proposal waits for review, and a human integrates it. A read-only token remains available if you want no writing at all.
Does Verbasil read my Claude conversations?
No. The MCP server receives only the questions the assistant asks it, never the rest of the exchange. The reading direction is one-way: Claude queries the memory, the memory does not crawl your conversations.
Do you need a different server for ChatGPT and for Claude?
No, that is the point of the protocol: the same MCP server serves ChatGPT, Claude, Cursor and Gemini. Do give each client its own token though: call limits are set per token, and two clients sharing a token also share its rate.