MMW Managed Memory Workspace

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](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 The dangerous failure is not forgetting — it is an agent confidently answering with an outdated fact.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 and a Telegram bot 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-0.1.14-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-0.1.14-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-0.1.14-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-0.1.14-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, search_history, forget, validate_memory, gateway_call, mmw_sync_status and spool_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-0.1.14-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 /