mneva-connect

Connect your coding agent to its hosted Mneva brain — persistent memory, belief revision, instinct, and the things no MCP memory server has shipped before: unprompted thoughts overnight, calibration from your own track record, allostatic awareness, and multi-agent outside views.

Every AI coding tool starts cold: it re-reads your codebase, re-asks what you told it last week, repeats the mistake you already corrected. Mneva is the mind underneath them — it remembers, it revises, it has instinct, and (the part that surprises everyone) it thinks while you sleep.

Setup

1. Get an API key at mneva.dev/signup.

2. Add this to your coding tool's .mcp.json (Claude Code, Cursor, or any MCP client):

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

3. Restart your coding tool.

What your agent gets

Memory and recall

| Tool | What it does | |------|--------------| | remember | Record a decision, a fix, a fact — kept across sessions | | recall | Semantic recall of past memories — finds them by meaning | | recall_dreams | Surface unprompted thoughts your brain produced overnight |

Belief, evidence, revision

| Tool | What it does | |------|--------------| | believe | Assert something held true about the project | | revise | Correct a belief — the old one is superseded for good | | evidence | Record one supporting/refuting observation; SPRT auto-adjusts confidence | | evidence_for | See the full evidence trail behind a belief's confidence |

Instinct and calibration

| Tool | What it does | |------|--------------| | flag | Mark code as fragile or dangerous | | assess | Gut-check before working — clear, caution, or danger, with your track record on adjacent predictions | | predict | Record a falsifiable claim so it can be checked later | | resolve | Close a prediction with what actually happened | | predictions_open | List unresolved predictions you owe answers on | | calibration | Your per-domain accuracy and confidence_gap — where to trust your gut, where to slow down |

Brain state

| Tool | What it does | |------|--------------| | state | Allostatic verdict — rest, engaged, or overloaded — and the numbers behind it |

Team brain (multi-agent)

| Tool | What it does | |------|--------------| | note_about | Leave an outside view about another agent sharing this brain | | outside_views | Recall observations agents have left about each other |

Environment

• MNEVA_KEY (required) — your tenant API key from the dashboard
• MNEVA_API (optional) — engine base URL, defaults to the hosted service
• MNEVA_AGENT (optional) — identify this agent for Team-plan multi-agent brains; defaults to default for solo use

What makes Mneva different

Mem0, Letta, and Zep all ship the same 5%: store memories, retrieve them semantically. Mneva's 95% is what came out of a year of work on a working cognitive substrate:

• Dreams — the engine wanders nightly, recombining your memories with a small LLM and surfacing the connections that crossed a novelty floor. No memory product on the market produces unprompted thoughts.
• Calibration in assess — the gut-check returns the specific past predictions you were wrong about in this domain, not just a verdict.
• SPRT belief revision — confidence rises and falls by sequential probability ratio with a queryable audit trail, not a hand-tuned heuristic.
• Allostatic state — the brain tells you when it is tired.

MIT licensed.