Claude memory, Projects, CLAUDE.md, MCP server: which memory for a small team?
On your own, Claude's built-in memory and a CLAUDE.md file are often enough. With a team, you need a memory that everyone reads and updates: shared Claude Projects (Team or Enterprise plans), a CLAUDE.md versioned in a repository for developers, or a shared MCP server you can use from claude.ai as well as Claude Code. mem0 and open source MCP servers are building blocks to assemble, mostly for individual use or for developers.
Your team has been using Claude for a few months. Everyone has their own habits: one person filled a Project with the client brief, another wrote a CLAUDE.md file in their repository, a third installed an MCP memory server found on GitHub. The day someone picks up a colleague’s file, nothing carries over: each person’s memory stays in its own corner. This article compares the options available in October 2026, based on their official documentation, to help you choose for your situation.
Six ways to give Claude a memory
Tools that call themselves Claude’s “memory” don’t keep the same things and aren’t built for the same people.
- Claude’s built-in memory. Claude draws memories from your conversations and saves them as it goes; you can also tell it “remember this.” Each Project has its own memory, kept apart from the rest. You can see and edit what it remembers. It is on by default on the Free, Pro and Max plans; on Team and Enterprise, organization owners decide whether to offer it and each member turns it on.
- Claude Projects. A Project groups documents, instructions and the conversations about one topic. Every account has access (up to five Projects on the free plan). Sharing with colleagues is limited to the Team and Enterprise plans, with two permission levels: view or edit.
- The CLAUDE.md file. A Markdown file that Claude Code reads at the start of every session. Placed at the root of a repository, it is shared with the team through version control (Git). Claude Code also keeps an “auto memory”: notes it writes itself, but that stay on the machine and are shared neither across computers nor between people.
- Open source MCP memory servers. The MCP project’s reference server stores a knowledge graph (entities, relations, observations) in a local file. Others, like MCP Memory Keeper, keep context in a SQLite database on your computer. They are designed for one person.
- Developer building blocks, like mem0. mem0 is a memory layer for AI applications, as a hosted service or a self-hosted version. It organizes memories by user, agent or session, and also offers an MCP server.
- A shared team MCP server. A hosted service, reachable over the internet, that each member connects their assistant to. The content (notes, decisions, sometimes tasks) is common to the whole team.
The comparison table
| Criterion | Claude’s built-in memory | Claude Projects | CLAUDE.md (and Claude Code auto memory) | Open source MCP memory servers (local) | mem0 and developer building blocks | Shared team MCP server |
|---|---|---|---|---|---|---|
| Shared between people | No: specific to each user | Yes, on Team and Enterprise, between organization members | CLAUDE.md: yes, through the Git repository; auto memory: no | No, usually single-user | Possible on the platform (organizations, projects, members), to set up yourself | Yes, that’s the whole point |
| Works in claude.ai | Yes (web, Desktop, mobile) | Yes | No | Not locally: claude.ai requires a server reachable over the internet; some offer a remote mode (MCP Memory Service) | Custom connector documented for Claude Desktop; claude.ai: to be checked | Yes, as a custom connector |
| Works in Claude Code | Not mentioned in the help center (Claude Code has its own memory) | To be checked | Yes | Yes | Yes (mem0 MCP server) | Yes |
| Updated by the assistant itself | Yes | To be checked | CLAUDE.md: if you ask it to; auto memory: yes | Yes, through the server’s tools | Yes (add, update, delete memories) | Yes, depending on the tools exposed |
| Tasks and tracking | No | No | No (free text) | Depends on the server (Memory Keeper: “task” categories, priorities) | No: memories, not tasks | Depends on the tool |
| Versions and history | To be checked | To be checked | CLAUDE.md: yes, through Git; auto memory: no | Depends on the server (Memory Keeper: checkpoints; reference server: no) | Yes: history of each memory (added, updated, deleted) | Depends on the tool |
| Access control | Organization settings; owners can’t see members’ memories | Per-member permissions: view or edit | Repository access; organization-wide CLAUDE.md possible | Whatever your computer provides | API keys, organization and project members | Accounts, OAuth, depending on the tool |
| Setup effort | None | Low | Low (one file), Git needed to share | Medium: local install and configuration | High: integration by developers | Low for users: add a connector |
A few notes on the shakiest cells. For Projects, Anthropic’s help center describes documents and instructions that you add yourself; it doesn’t say whether Claude can change that knowledge on its own, or whether files keep a history. It doesn’t mention using Projects from Claude Code either. For built-in memory, none of the pages consulted mention version history.
What each option does well, and where it stops
Built-in memory: great for you, invisible to others
It takes no effort and remembers your preferences, how you work, the context of your recurring topics. Its limit comes from what it is: it belongs to one person. If a colleague takes over your client while you’re on vacation, she gets nothing from what Claude learned with you.
Projects: the right habit of “one client, one Project”
One Project per client or engagement saves you from pasting the brief into every conversation. Shared on a Team or Enterprise plan, it gives the whole organization the same documents and instructions. Two limits to know: conversations stay private by default, even in a shared Project; and only members of the organization can be invited, not a client or an outside contractor. Finally, a decision made in a conversation only gets into the Project if someone adds it.
CLAUDE.md: simple and solid for developers
For a team working in a code repository, it’s often the best return on effort: a file reread every session, versioned, reviewed like the rest of the code. It stops where the repository stops. People who work in claude.ai without touching code have no access to it, and an agency with twenty clients ends up with twenty scattered files. Claude Code’s auto memory stays on each person’s computer.
Open source MCP servers: handy for one person
They give Claude Code or Claude Desktop tools to save and retrieve information, with no outside service. Installed locally, they only serve the person who installed them and can’t be reached from claude.ai. Some projects, like MCP Memory Service, offer a remote mode with OAuth and sync across devices: it can be done, but hosting and security are up to you.
mem0: a building block, not a turnkey team product
mem0 is designed for developers adding memory to their own application: memories per user or per agent, a history of every change, organizations and projects. Its MCP server lets you use it from a coding tool. For a small team without a developer, it’s a project rather than a tool.
The shared team MCP server: common content, several assistants
Once added as a custom connector, the same server works in claude.ai and, according to the Claude Code documentation, in Claude Code for accounts logged in with claude.ai. Everyone reads and adds to the same content. On the free plan, you can only add one custom connector; on Team and Enterprise, an owner adds it for the organization first. What varies from one tool to another: whether it has tasks, history, and how fine-grained permissions are. Check those three points before you choose.
Which memory for your situation
| Your situation | What’s usually enough |
|---|---|
| Solo, no code | Claude’s built-in memory and one Project per client |
| Solo developer | A CLAUDE.md per repository, Claude Code auto memory, possibly a local MCP memory server |
| A team, all developers, on the same repository | CLAUDE.md versioned in the repository, possibly a shared .mcp.json |
| A team, no code, everyone on a Team or Enterprise plan | Shared Projects, as long as decisions get added to the Project |
| A team with mixed profiles (claude.ai and Claude Code), several clients | Shared team MCP server |
| You’re building your own AI application | mem0 or a similar building block |
Three questions help you decide. Who needs to read this memory: only you, or also a colleague taking over the file? Where does the team work: in claude.ai, in Claude Code, or both? And does this memory need to track the work to be done too, or just the knowledge?
What about Tavimi?
Tavimi is the shared working memory for your team and your AI assistants, over MCP. It’s a hosted MCP server with OAuth sign-in that works with Claude Code, Claude on claude.ai, Claude Desktop and any MCP client. It belongs in the last column of the table, with clear answers to the three points to check:
- Tasks: status, priority, due date, owner (a person or a role like “copywriting”) and a dated log. When it’s missing a decision, the assistant asks the question on the task; only a human answers it.
- History: notes are versioned, nothing gets overwritten or deleted.
- Access: one space per client or project, an invited person only sees their space, two-factor authentication is mandatory, and when you connect an assistant you tick the spaces it can see.
Tavimi is neither project management software nor a personal memory for a single user: if you work alone, Claude’s built-in memory and a CLAUDE.md may be all you need. To see how a team uses it day to day, read How it works or the use cases.
FAQ
What people ask us
Is Claude's memory shared with my colleagues?
No. According to Anthropic's help center, memory drawn from conversations belongs to each user; even the owners of a Team or Enterprise organization cannot see or edit a member's memories. To share context, you need a shared Project, a common file or an external tool.
Can I connect an MCP server running on my computer to claude.ai?
Not directly. Claude connects to custom connectors from Anthropic's infrastructure, so the server must be reachable over the internet. A local MCP server works with Claude Code or Claude Desktop on the same machine, but not with claude.ai.
Is mem0 an alternative to Claude Projects?
Not really: mem0 is first a memory layer for applications that developers build, with an API and SDKs. Its MCP server also lets you use it from Claude Code or Claude Desktop, but it stores memories, not project documents or tasks.
Is a connector added in claude.ai available in Claude Code?
Yes, if you're logged in to Claude Code with a claude.ai account: the Claude Code documentation says connectors added in claude.ai are available there automatically. The same MCP server therefore works in both tools.
Sources
- Use Claude's chat search and memory to build on previous context (Anthropic), accessed October 3, 2026
- What are projects? (Anthropic), accessed October 3, 2026
- Manage project visibility and sharing (Anthropic), accessed October 3, 2026
- Get started with custom connectors using remote MCP (Anthropic), accessed October 3, 2026
- How Claude remembers your project (documentation Claude Code), accessed October 3, 2026
- Connect Claude Code to tools via MCP (documentation Claude Code), accessed October 3, 2026
- Knowledge Graph Memory Server (dépôt modelcontextprotocol/servers, GitHub), accessed October 3, 2026
- MCP Memory Service (dépôt doobidoo/mcp-memory-service, GitHub), accessed October 3, 2026
- MCP Memory Keeper (dépôt mkreyman/mcp-memory-keeper, GitHub), accessed October 3, 2026
- Mem0 MCP Server (documentation mem0), accessed October 3, 2026
- Memory History (documentation mem0), accessed October 3, 2026