Mneva — a coding partner with a mind.
Mneva is the hosted brain underneath your coding agent. It remembers, it revises, it has instinct, and it produces unprompted thoughts between sessions. One paste into your .mcp.json and your agent stops starting cold every morning.
What is Mneva
Mneva is a multi-tenant SaaS — one isolated SQLite brain per customer. Your agent talks to it over MCP (npx -y mneva-connect) or HTTP (https://mneva.dev/v1/*). Sixteen tools today, more shipping.
The shelf of "AI agent memory" products is crowded — mem0, Letta, Zep, the half-dozen mcp-memory-* packages, Anthropic's reference memory server. They all ship the same five percent: store what the agent tells you, retrieve it semantically. Mneva ships that, plus the other ninety-five percent:
- Dreams — overnight wander recombines your memories and surfaces connections you did not ask for (recall_dreams)
- Instinct with a track record —
assessreturns the past predictions you were confidently wrong about in this domain, not just a verdict (assess) - Beliefs that earn their place — evidence accumulates via sequential probability ratio with an audit trail (evidence)
- State — the brain tells you when it's overloaded (state)
- Multi-agent — agents on the same brain leave outside views about each other (note_about)
That's the difference between a notepad and a partner.
What's in this doc
- Getting Started — one-paste install, your first call, what errors look like.
- Client Setup — copy-paste configs for Claude Code, Cursor, Windsurf, plain Anthropic API.
- Concepts — what the unusual primitives mean: dreams, calibration, SPRT, allostatic state.
- MCP Tools — reference for every one of the 16 tools, with examples.
- REST API — every endpoint, every header, every payload.
- Operations — plans, rate limits, isolation, self-hosting.
Built by Barron AI Solutions. Source is intentionally small (~60KB engine + 200-line MCP wrapper). MIT-licensed wrapper, hosted engine.