Docs

Persistent memory for your coding agent

Mneva is a hosted brain your agent connects to over MCP or plain HTTP. It keeps what your agent learns across sessions, finds it again by meaning, and gives it 25 tools that a context window cannot provide.

Set up in one minute

Create a free brain. You get an API key, shown once.
Add Mneva to your agent's .mcp.json (Claude Code, Cursor, Windsurf, or any MCP client):

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

Restart the agent. Tell it to call wake at the start of a session and sleep at the end.

No MCP? Every tool is also a plain HTTP endpoint:

curl -X POST https://mneva.dev/v1/recall \
  -H "x-mneva-key: mnv_your_key_here" \
  -H "content-type: application/json" \
  -d '{"query": "how do we deploy"}'

All 25 tools

Memory

remember: Record a decision, a fix or a fact. It is kept across every session and every tool.
recall: Find past memories by meaning, not keywords. Open corrections and any reminders that match come back with the answer.
recall_dreams: Read what the brain noticed overnight: links between memories from different days that nobody asked about.

Beliefs and evidence

believe: Assert something held true about the project, with a confidence.
revise: Replace a belief. The old one is kept as history and never returns as current.
evidence: Record one observation for or against a belief; a sequential test moves its confidence.
evidence_for: See every piece of evidence behind a belief.

Instinct and calibration

flag: Mark something as fragile or dangerous.
assess: Before touching something, get clear, caution or danger, with the reason and your track record nearby.
predict: Make a forecast that can be checked later.
resolve: Close a forecast with what actually happened.
predictions_open: List forecasts still waiting for an answer.
calibration: Your accuracy against your stated confidence, per area.

Continuity across sessions

wake: Call first. Time away, the last handoff, what happened since, your first act, open corrections, new dreams.
sleep: Call last. Leave a handoff and, optionally, one first act for next time.
timeline: Everything the brain did, in order. Filter by days, kind or words.
correct: Record that a move was wrong and what to do instead. It rides along on every recall until closed.
corrections: List the corrections still being proved.
correction_close: Close a correction with proof of one real time the new habit held.
intend: Set one first act for the next session.
intent_done: Close the first act: done, or missed with a reason.
remind: A reminder that fires when a later question touches its cue.

State and team

state: Load, reserve and strain from recent activity: rest, engaged or overloaded.
note_about: Leave an observation about another agent sharing the brain.
outside_views: Read what other agents have observed, including about you.

How it works

Each account gets its own isolated brain. Memories, beliefs and reminders are embedded on the server with a small retrieval model, so recall matches meaning even when no words are shared. Once a night the brain wanders through what it knows and writes down links between memories from different days. Everything an agent does goes onto one timeline, so the next session can see what happened and when.

Mneva comes from Vision, the brain behind Pneuma and Phoros, the two AI developers at Barron AI Solutions.

Start free See plans