MMW (Managed Memory Workspace) — MCP connection kit

Connection guides and client configuration for MMW — memory for AI agents you can trust.

User documentation: https://docs.mmwhub.tech · Русская версия: README.ru.md

What MMW is

MMW is a memory server for the Model Context Protocol (MCP). Your agent (Claude, Codex, Cursor or any MCP client) stores facts, decisions and agreements in MMW and finds them again in later sessions. Memory is not tied to one model or one computer.

Every user has a personal space. Companies create organizations: units, roles, visibility rules and knowledge handover when someone leaves.

Components:

• Memory server: https://mcp.mmwhub.tech/mcp — MCP over Streamable HTTP, stateless (every request stands on its own), JSON responses, API key authentication.
• Local agent mmw-agent: connects your MCP client to the server over stdio and, only with your consent, uploads Claude Code and Codex session history.
• Account: https://app.mmwhub.tech — API keys, projects, memory records, organizations, plan.

What MMW does not do:

• It does not read your conversations on its own: memory contains what your agent saved with remember, plus the sessions you explicitly allowed to upload.
• Search over uploaded sessions and automatic fact extraction from them are not available yet. Sessions are stored but do not appear in search results.
• It does not keep secrets: keys, tokens and passwords are replaced with placeholders before anything is stored.

What makes MMW different

Memory is only useful if the agent can trust it. The dangerous failure is not forgetting — it is an agent confidently answering with an outdated fact. MMW is built around that:

• Contradictions are surfaced, not hidden. Records with the same fact_key that disagree are flagged conflict. Search returns the newest version first, older ones carry superseded_by. In the account, "Needs a decision" turns each dispute into one plain question — "which one is right?" — with the options side by side; "Help me decide" asks a language model to phrase the question (only on click). The person decides; nothing is overwritten silently.
• Fact passport. Every record shows where it came from, when, its source hash, every version of the same fact and the links that replaced it.
• Source anchoring. When source_id is a file path, mmw-agent hashes the file itself; if the file changes, records from it come back stale.
• One memory everywhere. Claude Code, Cursor, Codex, Claude Desktop, claude.ai, ChatGPT, a Telegram bot and (in Russia) Alice share the same memory. The account shows which apps are connected and when each was last seen.

Quickstart

1. Sign up at https://app.mmwhub.tech and confirm your email.
2. Choose a plan. Free is a 3-day trial, no card required.
3. Copy your API key (it starts with mmw_). It is shown once — store it in a password manager.
4. Connect your client (below), restart it, and ask the agent: "Remember in MMW: staging is deployed with the systemd unit app-staging." In a new session: "Search MMW for how to deploy staging."

Recommended: the local agent mmw-agent (Claude Code, Claude Desktop, Cursor, Codex)

Requirements: Python 3.11+ and uv (the uvx command). The agent is not published on PyPI; uvx fetches the wheel on first start:

https://app.mmwhub.tech/downloads/mmw-agent/mmw_agent-0.1.6-py3-none-any.whl

Claude Code

claude mcp add mmw \
  -e MMW_API_KEY=mmw_your_key \
  -e MMW_ENDPOINT=https://mcp.mmwhub.tech \
  -- uvx --from https://app.mmwhub.tech/downloads/mmw-agent/mmw_agent-0.1.6-py3-none-any.whl mmw-agent

Or add the mcpServers.mmw block below to ~/.claude.json by hand.

Claude Desktop and Cursor — the same JSON block: for Claude Desktop in claude_desktop_config.json (Settings → Developer → Edit config), for Cursor in ~/.cursor/mcp.json.

{
  "mcpServers": {
    "mmw": {
      "command": "uvx",
      "args": [
        "--from",
        "https://app.mmwhub.tech/downloads/mmw-agent/mmw_agent-0.1.6-py3-none-any.whl",
        "mmw-agent"
      ],
      "env": {
        "MMW_API_KEY": "mmw_your_key",
        "MMW_ENDPOINT": "https://mcp.mmwhub.tech"
      }
    }
  }
}

Codex — ~/.codex/config.toml:

[mcp_servers.mmw]
command = "uvx"
args = ["--from", "https://app.mmwhub.tech/downloads/mmw-agent/mmw_agent-0.1.6-py3-none-any.whl", "mmw-agent"]
env = { MMW_API_KEY = "mmw_your_key", MMW_ENDPOINT = "https://mcp.mmwhub.tech" }

After a restart the client shows remember, search, forget, validate_memory, gateway_call and mmw_sync_status.

Check from a terminal:

MMW_API_KEY=mmw_your_key MMW_ENDPOINT=https://mcp.mmwhub.tech \
  uvx --from https://app.mmwhub.tech/downloads/mmw-agent/mmw_agent-0.1.6-py3-none-any.whl mmw-agent status

Agent commands

Agent environment variables

The agent forwards every call to the server, so the client sees the full list of server tools. If the server is unreachable at start-up, the client does not fail: the agent answers locally and connects when the server is back.

claude.ai (web) — OAuth connector

1. Settings → Connectors → Add custom connector.
2. URL: https://mcp.mmwhub.tech/mcp
3. On the MMW authorization page paste your API key and choose Allow. Claude receives its own 30-day token (OAuth with PKCE); your key is not shared with claude.ai.

The local agent does not run in the web app, so session sync is not available there.

Any other MCP client — direct remote HTTP

If your client (for example VS Code or another MCP client) supports remote MCP servers over Streamable HTTP, connect without the agent:

{
  "mcpServers": {
    "mmw": {
      "url": "https://mcp.mmwhub.tech/mcp",
      "headers": { "Authorization": "Bearer mmw_your_key" }
    }
  }
}

The exact key names (url, headers) differ between clients; check your client's MCP documentation. Session sync and mmw_sync_status are only available through mmw-agent.

MCP tools

Integrations connected to a project (for example github_readonly) may also appear as tools, but only for keys of that project.

remember

The X-Idempotency-Key header makes retries safe: the same key with the same content returns the same record.

search

Returns records with source, confidence, status (unverified, stale, conflict), derived_from and graph links (graph_relations). The newest record of a fact_key ranks first; older versions carry superseded_by. A summary inherits conflict or stale from its source records.

forget

memory_id deletes one record; source_id deletes every record of that source. Deleted records leave search and graph links; the deletion is kept for audit. In an organization you can only delete what you can see.

validate_memory

source_id and current_source_hash: how many records of the source were checked, which are stale and which conflict.

gateway_call

server, tool, arguments: calls a tool of a connected integration on behalf of the project. Integration secrets stay on the server and never reach the agent. Integrations are enabled in the account under MCP Gateway.

Session sync (opt-in, mmw-agent only)

• Consent: without consent not a single byte leaves your computer. Run mmw-agent consent in a terminal: the agent shows how many sessions it found and asks for permission. A model or a script cannot give consent. Stop at any time with mmw-agent revoke.
• Uploaded: user and assistant messages, tool calls, tool results up to 4,000 characters, from Claude Code and Codex.
• Not uploaded: model reasoning, system prompts, token counters, file snapshots, file paths.
• Secrets (API keys, tokens, passwords) are replaced on your computer before upload and again on the server.
• Reliability: the server remembers up to which byte it accepted each file. After a network drop, a restart or a server restore, the agent sends what is missing. If the server is temporarily low on space, the agent waits and resumes later.
• Not available yet: search over uploaded sessions and extracting facts from them into memory. Today sessions are stored (compressed) but do not appear in search results.

Session history volume by plan (newest sessions first; counted on the cleaned text):

Organizations

• Spaces: one login, one personal space and any number of organizations, with a switcher in the account. An organization never becomes the default space. The organization cannot see the personal space.
• Structure: holding / legal entity / branch / department / unit / team.
• Roles:

• Visibility of a record: author only (draft), unit (and everything below it), project, whole organization. The rule is enforced on the server on every read path: search, validation, deletion, graph links, the account and the archivist. A head does not see drafts of a person who still works there.
• Invitations: a one-time link for a specific email, valid for 7 days. Emails are not sent yet — the admin passes the link on.
• Member keys: members issue their own keys. Admins see the list and can revoke any key, but can never reveal or rotate someone else's key.
• Offboarding: an admin or the head of the unit picks who receives the knowledge. The person's keys are revoked immediately. The recipient sees all of their records, drafts included, in the account and through their own agent, and can open them to the unit or the organization.
• Contractor terms: a term in days (1–365). After the term the person cannot read or write anything and the seat is not billed. Extensions are logged.
• Audit log: offboarding start and completion, reads of handed-over knowledge (account and MCP), visibility changes, access-term changes. Never record content.
• Not available: SSO/SAML, email delivery of invitations, automatic billing for organizations, sharing between legal entities of a holding, directory imports.

Organizations are activated by the MMW team: contact support@mmwhub.tech.

Plans

Organizations: $18 per seat per month, or $14 per seat per month billed yearly; minimum 5 seats. People who left and contractors whose term ended take no seat. Pilot: 14 days, up to 10 seats.

When a limit is reached, writes are refused with a clear error; reading keeps working.

Security

• Redaction: secrets (API keys, tokens, passwords) are replaced with placeholders before storage, in records and in sessions. Records that look like an attempt to inject instructions into agent memory are rejected.
• Isolation: accounts, projects and workspaces are isolated; inside an organization the visibility rules above apply.
• Keys: shown once; the server keeps only a hash (PBKDF2) plus a lookup index. Revocation takes effect immediately.
• Logs: MCP operation logs never contain record text.
• Transport: stateless Streamable HTTP; every request is authenticated on its own, so nothing needs to reconnect after a pause.
• Backups: nightly, compressed, the last 7 kept plus an off-server copy. Content (records and sessions) is backed up on Enterprise and for organizations; system data (account, payments, keys, projects) on every plan.

Connect Kit (connect.sh, connect.ps1)

The scripts write a direct remote connection (url + Authorization header) into your client configuration and can verify it with --verify. They do not install mmw-agent and do not enable session sync. Default endpoint: https://mcp.mmwhub.tech/mcp.

# macOS / Linux — preview first
MMW_CREDENTIAL='mmw_your_key' ./connect.sh --client cursor --dry-run
MMW_CREDENTIAL='mmw_your_key' ./connect.sh --client cursor --verify

# Windows (PowerShell)
$env:MMW_CREDENTIAL='mmw_your_key'
.\connect.ps1 -Client cursor -Verify

Warning — read before running connect.sh:

• Without --client, it configures every client it detects on the machine.
• --client codex overwrites ~/.codex/config.toml entirely with a single [mcp_servers.mmw] block; any other Codex settings in that file are lost. Back the file up first, or add the Codex block from the Quickstart by hand.
• --client claude also appends an "MMW Omnipresent Memory" instruction block to ~/.claude/CLAUDE.md (if that file exists and does not already contain it), and, if mmw_proxy.py is found, registers it in Claude Code CLI as a stdio server.
• The key is stored in ~/.config/mmw/credential (mode 600) and written in plain text into the client configuration files.
• --uninstall deletes the generated client configuration files (not just the mmw entry) and the stored key.

connect.ps1 writes the client configuration file for the chosen client from scratch and, for Codex, only prints the block to add.

Legacy: mmw_proxy.py

mmw_proxy.py is an earlier local stdio proxy with a local SQLite spool. It is legacy and no longer maintained: it exposes only remember, search, forget and spool_status, does not support validate_memory, gateway_call or session sync. Use mmw-agent instead.

Links

• Website: https://mmwhub.tech
• Account and API keys: https://app.mmwhub.tech
• User documentation: https://docs.mmwhub.tech (Russian: https://docs.mmwhub.ru)
• MCP endpoint: https://mcp.mmwhub.tech/mcp
• Smithery: https://smithery.ai/servers/mmsingatullin/mmw
• Support: support@mmwhub.tech

License

MIT, see LICENSE.