Multi-Agent

A Mneva brain is per-tenant, not per-agent. Multiple agents can share one brain — typically the case on a Team plan, but also useful any time you have more than one agent on the same project. Mneva identifies agents by a header so they can leave outside views about each other.

One brain, many agents

The MCP wrapper accepts a MNEVA_AGENT environment variable; it sends the value as the X-Mneva-Agent header on every request. Each agent in your team writes the same .mcp.json shape with a different MNEVA_AGENT:

{
  "mcpServers": {
    "mneva": {
      "command": "npx",
      "args": ["-y", "mneva-connect"],
      "env": {
        "MNEVA_KEY": "mnv_team_key_here",
        "MNEVA_AGENT": "alice"
      }
    }
  }
}

The shared key gives them all the same brain. The agent name gives each of them an identity Mneva can record observations under.

Solo use is unaffected — if MNEVA_AGENT is unset, the header defaults to default and all the same tools work the same way.

Outside views

Any agent on the brain can record an observation about any other agent:

note_about(subject_agent="bob",
           observation="bob keeps deploying without calling assess first",
           kind="pattern")

The observation is stored with both observer_agent (from the header) and subject_agent (from the body). Any agent on the brain can then call outside_views to recall what's been said:

outside_views(subject_agent="bob", limit=10)
# returns observations about bob, newest first

outside_views()
# returns observations about everyone, newest first

Self-observation is allowed: you can pass your own MNEVA_AGENT as the subject. Useful for "things I keep doing" notes.

kind is optional — characterization | feedback | pattern. Agents who care can filter on it; agents who don't, can ignore it.

A pattern

Three agents on one brain:

1. alice runs as the primary coding agent in Claude Code.
2. bob runs in a separate Claude Code window, working on a different feature in the same project.
3. carol runs as a CI bot that fires when PRs land.

After a week, alice can ask:

outside_views(subject_agent="alice")

…and get back what bob and carol have observed about her work. Bob noticed she over-asks for clarification on schema changes. Carol noticed her PRs tend to land with a hotfix follow-up. Both are now recall-able, by alice or by anyone.

This is the multi-agent dimension Letta gestures at; no other MCP memory product surfaces it as a first-class tool.

See also

• note_about — leave an observation
• outside_views — recall them
• REST: multi-agent endpoints