# Baron Munchausen — memory that has to back every claim

> A local memory server for AI agents. The agent writes an answer; the server
> checks it sentence by sentence against your own graph and returns a verdict —
> grounded, partial or ungrounded — with the sentences nothing backs. Python
> 3.12, standard library only, MCP over stdio and JSON-RPC on loopback.
> Apache-2.0, self-hosted free. One paid operation: the hosted verdict.

Measured 8 September 2026. Every number below is produced by the repository or
by a live probe; nothing here is rounded up for effect.

## Start here

- [Install](docs.html#install): three commands, two minutes, no dependency to install first.
- [Clients](docs.html#clients): Claude Code, Codex, Claude Desktop, Cursor, llama.cpp, LangChain.
- [JSON-RPC reference](docs.html#rpc): two routes — `GET /health`, `POST /` — and 26 tools.
- [Catalogue](catalog.html): 1507 public API records shipped in the graph, with a live status.
- [Pricing](pricing.html), [Console](console.html), [Hub](hub.html), [Security](security.html).

## The contract we ask an agent to follow

1. `memory_ground_prepare` (or `memory_checkpoint`) before answering.
2. Answer using only what came back.
3. `memory_ground` after answering — the verdict, and the claims nothing backs.
4. `memory_add` for anything worth keeping.

Skipping step 1 makes the verdict `ungrounded` by construction, however well the
answer reads. That is deliberate: memory cannot be skipped quietly.

## Thresholds

- A claim is supported at coverage 0.60 and partial from 0.30.
- An answer is grounded at 0.80 of claims supported with none unsupported, and partial from 0.40.
- Every node passes P1–P6 before storage — freshness, source, numbers, consistency, reproducibility, completeness — and is stored at four of six or better.

## Prices

- The engine, all 26 tools, all reads and writes, self-hosted: **free**, Apache-2.0.
- Hosted `memory_ground`: **0.0035 USDC per call**, USDC on Base (`eip155:8453`) through an x402 challenge. Provisional until the cost of producing a verdict is measured.
- Free verdict endpoint, no key and no account: 20 a day per address, 200 with a key.
- Priced and not built: delta analyser $12/month or $0.50/day; product map $1 up to 200 files, $3 up to 1000.

Withdrawn, and not to be quoted from anywhere: 0.0035 USDC per memory
operation, 0.005 USDC per query, the six subscription tiers (Free, Solo $19,
Pro $49, Team $149, Scale $499, Enterprise), annual discounts, prorated
upgrades, prepaid credits of any kind, and a founder programme.

## Honest numbers

- Revenue, lifetime, both services: $0.031. Payers today: 0.
- Tests: 973 passing. Tools: 26. Catalogue: 1507 records — 1229 alive, 70 needing a key, 110 dead, 98 not confirmed.
- The repository is not public yet. Every registry listing waits behind that.

## Paid routes for agents

Eighteen routes on a sibling service answer an HTTP 402 challenge — USDC on
Base, no key, no account. Every paid route has a free `/preview` returning the
real shape of the answer. Listed on the [catalogue page](catalog.html#x402).

## Contact

hello@shinegang.click — read by a person.