Give your agents a read-only window into your memory.

// what it is

A broker, not a back door.

When an AI agent on your Mac wants to look something up in your history, it can't touch your recordings or database directly. Instead it asks the mnema command-line tool, which checks that you've approved that tool, that the request is within the scope you approved for it, and that the results are redacted — then hands back only what's allowed. You stay in control of who gets access and how far back it reaches.

Under the hood, every request is mediated by Mnema's broker: the same privacy policy, redaction, and retention rules that govern the app apply to agent access. Tools never read the database or media files directly — every call goes through the mnema CLI, which talks to the running app over a local, per-user channel. Access is read-only and can be blocked from Settings at any time.

// guarantees

What an agent can and can't do.

• ✓
  Read-only Agents can search and read your history. They can't edit, delete, or change anything.
• ✓
  You approve each tool Every client is approved individually, from inside Mnema. Nothing gets in without your explicit grant.
• ✓
  You set how far back — and it doesn't expire You choose how far back a tool can read — the last day, the last 7 days, or all retained history. Access has no clock on it: it stands until you block it, and lapses on its own 30 days after the tool last used it.
• ✓
  Redacted at the source Results come back already redacted by the same rules the app uses — passwords, keys, and tokens are stripped before an agent ever sees them.
• ✓
  Blockable any time Block a tool from Settings → Access and it is refused immediately — with no approval window — until you re-enable it there.
• ✓
  Brokered, never direct The skill always goes through the mnema CLI — never a direct database or media read.

// setup

Two steps to wire it up.

1. 01

   Install the mnema CLI

   In Mnema, open Settings → Access and install the CLI. Mnema places the mnema command in its managed CLI directory — usually ~/.local/bin on macOS. If that directory isn't on your shell's PATH yet, Settings shows you the one-line follow-up to add it. Access is read-only, scoped, and approved per client from inside the app.

2. 02

   Add the mnema-data skill to your agent

   In Claude Code, install it from the Mnema marketplace:

   claude code

   1. > /plugin marketplace add shaik-zeeshan/mnema
   2. > /plugin install mnema-data@mnema
   3. > /reload-plugins

   Using Codex, Cursor, OpenCode, or PI? Install it in one command with the open-source skills CLI:

   terminal

   1. $ npx skills@latest add shaik-zeeshan/mnema

   The skill always goes through the CLI, never a direct database or media read.

// usage

Then ask your agent for things like…

You won't usually type these yourself — your agent runs them for you. But here's the shape of what's happening under the hood.

access request --scope last-day

Optional. Asks you to approve read-only access covering the last day of history, before an agent needs it — you confirm inside Mnema. Agents don't need to run it: any data command opens the same approval window by itself.

search --query "launch notes" --format json

Searches your redacted history and returns structured results — timestamps, source app, matched text — ready for the agent to reason over.

show-text <result-id>

Pulls the full redacted text for one result, so the agent can read it in context.

open <result-id>

Opens that exact moment in the Mnema app — the full-res frame or audio clip — so you can verify it yourself.

// chat apps

No terminal? Ask from a chat app.

The same brokered access works from desktop chat apps that speak MCP — like the Claude desktop app. Run mnema mcp and the CLI becomes a local MCP server exposing the same five read-only tools: search, timeline, speakers, full text, and open-in-Mnema. Same approval dialog, same grants, same redaction — nothing new to trust.

Add this to your chat app's MCP configuration (for Claude Desktop: Settings → Developer → Edit Config), replacing <you> with your macOS username:

{
  "mcpServers": {
    "mnema": {
      "command": "/Users/<you>/.local/bin/mnema",
      "args": ["--client", "Claude Desktop", "mcp"]
    }
  }
}

The first question that needs your history pauses while Mnema shows its approval dialog — approve once and it stays approved until you block the tool.

// the other direction

Bring your tools into Ask AI.

This page is about letting outside agents read your history. The reverse works too: inside the app, Ask AI can connect to MCP servers you configure, so its answers can call your own tools alongside your history. Set it up in Settings → Intelligence — nothing connects unless you add it.