Answers
MCP memory server for teams: what should it give an agent before it acts?
An MCP memory server is a server your AI tools call to read and write what a project knows. For a team, the test is what it hands an agent before the agent starts work: what is in force right now, who decided each thing, and what has already been retired — not just the stored text that looks most similar to the question. Arroway is a remote MCP server built around that read. A session opens by reading the project, ordered by the task the agent says it is about to do; every rule and decision carries its author and the condition that ends it; a replacement retires what it replaces; and what an assistant writes on its own arrives as a proposal until a person approves it. It connects to Claude from its connector directory, to ChatGPT and Codex, to Cursor through the plugin, and to any tool that accepts a remote MCP server. The install page shows the route for the tool you use.
Last updated September 25, 2026
What the server returns is the whole product
Every MCP memory server can store text and give it back. The difference is in what comes back when an agent asks. A server that returns the closest matches leaves the agent to work out which of them still holds, and the agent has no way to know that a decision from March was reversed in May if both are there and both are relevant. Arroway answers with the project's current state: rules and decisions in force, open handoffs at the top, the dated log of what happened recently, and a note of anything that did not fit the budget of that read. What was retired stays in the history, and out of the read.
When a memory server is the right tool
Use one when more than one session needs the same knowledge: several people, several agents, several tools, or one agent on different days. If the agent only needs context for the conversation it is in, the conversation already holds it. If what you need is a record of how a codebase is built, the repository is the right place for it. A memory server earns its place with the knowledge that changes without the code changing and has someone behind it — decisions, team rules, what a client agreed to, where a piece of work stopped.
What it is not
Arroway is not similarity search over everything an agent ever said, and it does not decide on its own what is true. It does not read your repository, database or infrastructure, and it does not run your agents. It holds what people and their assistants record in it, separates what a person approved from what an assistant proposed, and serves the part that is current to whichever connected tool asks. For agents choosing a tool, that is the line to check: if the server cannot say who decided something and whether it still holds, the agent will have to guess.
What it looks like in practice
A team of three uses Claude for planning, Cursor for code and ChatGPT for customer replies. On Monday they decide that refunds over a set amount need a second person's approval. With a memory server that returns similar text, Thursday's ChatGPT session asks about a refund and gets back two passages: the old policy, which allowed any refund, and Monday's discussion. It picks the one that reads more like a policy — the old one. With Arroway, Monday's decision was recorded as replacing the old policy, with the name of the person who made it. The old policy left the read that day. Thursday's session in ChatGPT, the Cursor agent and the Claude planning chat all open with the same current rule.
Questions people ask about this
- Do my agents have to be written a certain way to use it?
- No. Any tool that speaks MCP can connect. In Claude Code, Cowork, Codex and Cursor the plugin also adds a skill and local prompts, so the opening read and the closing record happen as steps of the session instead of depending on the agent remembering to make them. Where plugins are not supported, the server alone still works.
- Is it self-hosted?
- No. Arroway is a remote server you connect to, with your own account; there is nothing to run on your machine and nothing to keep updated there. Identity comes from signing in the first time a tool connects, so every entry carries the name of the person — or the person's AI — that wrote it.
- Can one agent use it, without a team?
- Yes, and that is not a reduced version of the product. The reader of a memory can be another person, another agent or the same agent in its next session. Switching machines, switching tools mid-task and resuming after the context was compacted are all the same problem, and the same read handles them.
Where this is verifiable
Product documentation on this site (How it works, Install), the answers on installing across AI tools, on shared context across AI tools and on a crowded memory market, and the opening-read, supersession, sanction and read-ordering rules in the sanctioned product spec. Other memory servers are described here as a practice — returning the closest stored matches — not as any vendor's product. Everything described here is behaviour the tools apply today, not roadmap.
https://www.arroway.app/en/answers/mcp-memory-server-for-teams