Let an AI agent use your memory

Read this if you use an AI agent and want it to automatically reach into your work context — without you pasting in emails or saying “go check OMem.” This is the payoff: the agent answers from your real history on its own.

What it looks like

Once connected, the agent recognizes a work-flavored question, queries OMem, reads just what it needs, and answers — citing your actual emails, docs, and notes. Step through a real exchange:

You

What were the open items we promised to revisit from the Acme Q3 review?

A question about your real work — the kind an agent normally has no context for.

Next stepturn 1 / 6

The key: you don’t tell it to use OMem. It notices the question touches your work and reaches for it.

Three ways to connect

Pick the one that matches your agent.

Skill (recommended)

The richest integration, for agents with a skill / plugin mechanism — Claude Code, Codex, Claw Hub. The skill carries trigger semantics and the drill-down flow, not just raw tools. First make sure the omem CLI is installed and set up (install → setup).

In Claude Code:

/plugin marketplace add seacen/omem
/plugin install omem@omem-marketplace

In Codex and Claw Hub, install OMem’s plugin from their respective plugin systems — the same skill, a different front door. (Full per-agent steps.)

To verify, ask the agent “check my OMem health” — it should run omem doctor.

MCP (fallback)

For agents without a native skill mechanism (Cursor, Cline, Continue, Zed), OMem’s CLI runs a stdio MCP server (omem mcp) exposing the same six tools. One config entry connects it:

Cursor — .cursor/mcp.json (or Settings → MCP Servers):

{
  "mcpServers": {
    "omem": { "command": "omem", "args": ["mcp"] }
  }
}

Cline, Continue, Zed — same shape: command: "omem", args: ["mcp"], in that agent’s MCP config. (If your agent does support a skill, prefer that — you get the trigger + drill guidance the raw MCP entry doesn’t carry.)

Plain CLI

No integration needed at all. Any agent that can run a shell command can use OMem — it’s just a CLI:

omem query "what did Alice say about Q3 scope" --format json

The skill and MCP server are thin wrappers over exactly this. If your agent can bash, it can already query OMem.

What the agent can — and can’t — do

Both the skill and the MCP server expose the same six read-only tools:

Tool	What it does
query	Ranked search across all your indexed work
page_get	Full wiki page (frontmatter + body)
raw_get_path	Path to the original source file (file kind)
raw_get_parsed_path	Path to the parsed source (with version history)
wiki_ls	Browse pages by kind / source / account / date
doctor	OMem health check

The agent can never change your data

Setup, ingest, re-indexing, purging — none of these are exposed to the agent. Those stay deliberate, human actions you run in the terminal. The agent only ever reads, so handing it OMem access can’t reconfigure or delete anything. See querying from an agent for how a well-behaved agent uses these tools.

The CLI is the real interface — everything else is a thin wrapper

The skill (the plugin you install in Claude Code) is a small wrapper that shells out to omem query. The MCP server is a thin process that does the same. Neither re-implements anything — a human in the terminal and an agent through the plugin run the identical command and get identical behavior. That’s deliberate: the agent layer changes year to year, so OMem keeps the memory underneath bound to a stable CLI, not to any one agent. Switch agents whenever; your memory persists. (Design principle P5.)

What’s next

You’ve gone from nothing to an agent that remembers your work. From here:

• How-to guides — ingest more sources, tune search, schedule it, back it up.
• Concepts — why OMem is shaped the way it is.