Knotenhirn
Sign in Try it free

All guides

Team memory via MCP: what it is and when it pays off

Updated 15 November 2026 · by Michael Schwarz

A team memory is a shared place for short notes that AI agents search and add to via MCP. It pays off as soon as several people work with agents every day and keep explaining the same things: edge cases, agreements and the why behind decisions. It replaces neither the wiki nor the rules file in the repository. It fills the gap between them.

What a team memory is

An AI agent starts every session without a memory. What it learned yesterday it no longer knows today, and what it learned with a colleague it never knows with you. A team memory changes that: it is a store of short notes that every agent in the team reaches via MCP. MCP (Model Context Protocol) is the open standard through which agents such as Claude Code, Claude Desktop or Codex use tools and data.

The agent searches there when a task calls for it and gets only the notes that fit the question. When it learns something new in a conversation, it records it as a note. From then on it is available to everyone, to the people in the browser and to the agents of the others.

Where it sits between the rules file and the wiki

Most teams already have two places for knowledge. A team memory is the third, and each has its job:

  • The rules file in the repository, for example a CLAUDE.md, holds what applies to every task: commands, conventions, things not to do. It is always loaded and therefore has to stay short.
  • The wiki holds what people read from start to finish: requirements, concepts, handbooks. The pages are long and stay valid for a while.
  • The team memory holds what is needed only sometimes, but then urgently: edge cases, agreements with customers, pitfalls and reasons for decisions. It is not loaded, it is queried.

More on the line between file and memory is in the guide CLAUDE.md or knowledge base, more on the line to the wiki under Wiki and tacit knowledge.

How you can tell you need one

  • You are explaining the same thing to your agent for the third time.
  • Two people get different answers from the agent to the same question, because each told it something different.
  • The rules file is growing and now contains things that apply to one module or one customer only.
  • A finding from a long debugging session lives only in a closed ticket.
  • New team members ask five people before they have an answer.

If only one person works with one agent on one repository, the rules file is usually enough. The benefit of a shared memory grows with the number of people, repositories and agents.

What belongs in it and what does not

Good notes can be understood on their own: a title, a few sentences, a reason. For example: the import for a certain customer runs at night because their system is locked during the day. Or: an interface returns amounts in cents although the field is named differently.

Whole documents, credentials and anything that can be read from the code do not belong in it. Long texts stay in the wiki. The memory records at most what follows from them and points to the page.

How it stays current without creating work

A wiki goes stale because someone has to find time to maintain it. A team memory turns that around. The agent writes by itself while it works, and a person only decides what stays. Three things help:

  • When saving, the agent sees similar notes and merges duplicates.
  • Many notes at once, for example from existing documentation, arrive as proposals. A person accepts, edits or rejects them.
  • Searches without a result are collected. That shows you which knowledge is missing before someone asks for it.

What to look for when choosing

  • Permissions: does the agent see only what the person who connected it may see?
  • People: can everyone in the team read and change the notes in the browser, or is it an interface for developers only?
  • Vendor: does it work with every agent that speaks MCP, or with one only?
  • Location: where is the data, where does the AI processing run, and is there a data processing agreement?
  • Cost: is there a fixed price, or is it billed by calls and tokens?

What this looks like with Knotenhirn

Knotenhirn is a hosted team memory for AI agents with a built-in MCP server. People write notes in the browser, agents read and write them via MCP, and groups define who sees what. Notes organise themselves into topics, and the knowledge gaps page shows searches without a result. The servers are in Germany, the AI processing runs in the EU. How Knotenhirn differs from other tools is shown in the comparison.

Try it with your team

Try Knotenhirn free for 14 days, with no payment details. The trial does not turn into a subscription automatically.

Try free for 14 days

Knotenhirn

The shared memory for your team and your AI agents. Complements your wiki.

Servers in Germany · AI in the EU

Try free for 14 days

Product

  • Features
  • Pricing
  • Comparisons
  • Connect MCP
  • Guides
  • Sign in

Security

  • Security and DPA
  • DPA as PDF
  • Privacy

Legal

  • Legal notice
  • Terms
  • Right of withdrawal

Contact

  • knotenhirn@michael-schwarz.dev
  • 15 minutes with Michael
  • LinkedIn
  • GitHub
  • michael-schwarz.dev
© 2026 Michael Schwarz Cancel contracts here Deutsch