MemoryIntelligence MCP Server

Stop paying AI to reread the same context.

Receipted memory for your AI, via MCP. What you tell your assistant becomes structured memory you own — recalled by meaning, with every answer cited to its source. Works with Claude Desktop, Claude Code, Cursor, VS Code, and any MCP client.

Start in 30 seconds

pip install memoryintelligence-mcp     # or: pipx / uvx / uv tool install
mi-mcp setup                           # paste your key once — wires everything
# restart your assistant, then just talk to it:
#   "remember we picked Postgres for billing — we needed transactions"
#   (new session)  "what did we decide about the billing database?"

mi-mcp setup stores your key securely (macOS Keychain, or a chmod 600 keyfile), wires your assistants, opts the current folder in for capture, and verifies it — in one command. Your API key is never written into a config file.

👉 Get a free API key · Product · Issues

What you get

Three tools, ready the moment it's wired — a compatible host (Claude Desktop, Claude Code, Cursor) recalls and captures on its own, no prompts to memorize:

And four things that make it more than a notepad:

• Receipted — every recall cites the memory it came from. It cites, it doesn't guess.
• Reusable — capture once, recall by meaning across every session and every tool.
• Owned — memories are portable structured objects in your account, not locked in a model.
• Private — capture is opt-in per project; PII is redacted from what the agent sees.

The default surface is 7 tools (also mi_upload, mi_verify, mi_forget, mi_workspaces). Set MI_MCP_FULL=1 for the full 11-tool surface (adds mi_batch, mi_explain, mi_match, mi_account). Tools outside the active surface are rejected at the call boundary, not just hidden.

Capturing into a shared workspace

By default every capture is personal — it lands in your home workspace and nobody else sees it. To put one in a team space, name it:

"List my workspaces." → mi_workspaces "Save that to Somewhere Inc." → mi_capture(workspace_id=…)

Two rules keep a shared space from filling up by accident:

• A workspace with more than one member needs an explicit confirm. The first call returns a preview — "this posts to Somewhere Inc, visible to 4 members" — and saves nothing. Only after you say yes does the agent re-call with confirm=true. A fresh chat can't inherit that approval.
• Every capture tells you where it went, and a workspace you don't belong to is refused rather than quietly saved somewhere else.

Reading works the same way — mi_ask and mi_list take the same workspace_id:

"What did the team decide about billing?" → mi_ask(workspace_id=…)

Omit it and you search your own memories. One caveat worth knowing: whether a workspace read returns other members' memories is a server-side setting that is off by default. The result's scope block reports member_wide_reads so the assistant can tell you what it actually searched instead of assuming.

How it works

You ──"Remember we picked Postgres for billing — we needed transactions."──┐
                                                                mi_capture  ▼
                    ┌──────────────────────────────────────────────────────┐
                    │  MemoryIntelligence  (your account, over HTTPS)        │
                    │  → a structured, searchable, provenanced memory —      │
                    │    owned by you                                        │
                    └──────────────────────────────────────────────────────┘
                                                                    mi_ask  ▲
You ──"What database did we choose for billing, and why?"───────────────────┘
   ◀── "Postgres — you needed transactions."   (cites the memory it came from)

The server is a thin local layer: an MCP tool call becomes an authenticated HTTPS request to your MemoryIntelligence account. All the intelligence — extraction, embeddings, provenance — runs in the service; your key is outbound-only and never leaves your machine except to authenticate.

# macOS — Keychain:
read -s K; security add-generic-password -a "$USER" -s "MI_API_KEY" -w "$K" -U; unset K

# Linux / Windows — chmod 600 keyfile:
mkdir -p ~/.memoryintelligence
umask 077 && printf 'MI_API_KEY="%s"\n' "$YOUR_KEY" > ~/.memoryintelligence/.env

# then, on any OS:
mi-mcp wire
echo "$(pwd)" >> ~/.memoryintelligence/mcp/opt-in-paths   # allow captures here

mi-mcp doctor           # checks version, binary, PATH, key, wiring, opt-in, vault path
mi-mcp status           # wired surfaces + opt-in allowlist
mi-mcp wire --dry-run   # preview wiring changes

[✗] version  0.2.5 installed, 0.2.8 available — `uv tool upgrade memoryintelligence-mcp && mi-mcp wire`

{ "servers": { "mi-local": { "type": "stdio", "command": "mi-mcp" } } }

pip install -e ".[dev]"          # from mcp-server/
PYTHONPATH=src python -m pytest
ruff check src/

Learn more: memoryintelligence.io · Get a key · API reference · What is MCP · Changelog

Apache-2.0 © Somewhere Media, LLC. See LICENSE.