The world calls him a liar. He backs every word.
That is the whole product in a sentence, and it is not a
mood. memory_ground reads a finished answer, checks each
sentence against the graph, and returns grounded, partial or ungrounded
with the coverage number attached.
The company is shinegang. The memory
product is Baron Munchausen, short command name
baron. Two levels, on purpose: a company outlives the name
of any one thing it makes.
Why the name changed
- The old one was taken
mnemosis already on PyPI, and by the same class of product: “Memory for agentic AI”, and a second package, “Standalone memory & knowledge server for AI agents”. Measured 8 September 2026. We would have had to move even without a rebrand.- The new one is the mechanism
- The Baron is the man nobody believes who turns out to be right, and proves it. That is literally what the paid operation does. A name that describes the mechanism does not need a tagline explaining it.
- It is free in every registry we need
baron-munchausenis the one spelling that is available on PyPI, npm and the MCP registries at once, so the package, the command and the listing can all agree.
Where we stand against the field
Four honest paragraphs would be shorter than four comparison pages, which is what the old site had. The distribution argument is settled and not in our favour: Mem0 has roughly 64,000 stars, an SDK ecosystem and a place in other people’s agent kits; we have none of that. What follows is the part that is actually different.
| Against | Their strength, stated plainly | What we do instead |
|---|---|---|
| Mem0~64k stars, Apache-2.0, managed cloud | Distribution and ecosystem, and it is not close. Graph mode exists as a paid tier. | No verification happens at write time there: whatever extraction produces is what the memory holds. Here every fact passes six checks before it is stored, and the graph is the storage model on every install. |
| Cognee~30k stars, Apache-2.0, ECL pipeline | Ontology validation: a fact must fit the declared schema, which keeps traversal predictable. | Form and content are different questions. A statement can be perfectly schema-valid and completely wrong. We validate content, and we chose that because it is the harder one. |
| Zep / Graphiti~30k stars, Apache-2.0 | Bitemporal edges — the most rigorous time model in agent memory. It can reconstruct what the system believed last Tuesday. We cannot. | We try to keep the bad fact out rather than record when it was believed. And one SQLite file against Graphiti plus Neo4j or FalkorDB is a different operational bill. |
| All of them | Every one is more adopted, more tested in production, and easier to hire for. | Not one of them returns a number saying how much of an answer its own memory backs. That is the only thing here worth charging for, and it is the only thing we charge for. |
Where we fall short, in one line: adoption. Stars are not a quality measure, but production hours are, and they have far more of them.
The state of things, without the polish
| Revenue, lifetime | $0.031. Sixteen settled payments, two external buyers. |
|---|---|
| Payers today | 0. |
| Repository | Not public yet. Every registry listing waits behind that one action. |
| Search engines | We do not appear. Not the site, not the repository, not the profile. |
| Audience | One channel with a pulse: 28 followers on a developer forum, four to seven votes a post. |
| Engine | 973 tests passing, 26 tools, 1507 API records shipped in the graph. |
Publishing this table is the same discipline as publishing a verdict. A page that grades other people’s answers has to show its own numbers, including the ones that are bad.