Product · in development
Evidence first.
Then everything around it.
Hadano AI Cabinet is a local, evidence-first knowledge database for AI agents: it finds the passage that backs an answer, then expands around it. Your knowledge lives in a single file on your own machine, searchable over the Model Context Protocol. No server, no API key, no GPU.
Status
Not purchasable today, and we would rather say so
The package is not on PyPI and the repository is not public, so pip install does not work. Everything on this page describes what has been built and measured, not what you can download. Seven of nine specified MCP tools are implemented and reachable over stdio; two are designed and not built, and calling those fails rather than returning something plausible.
The release is waiting on multi-platform builds. The storage engine is C and needs a build per platform; it has been built and run on Windows x64 so far, with macOS, Linux and ARM next.
01 / Mechanism
Search in two stages, not one
Most AI memory tools run a semantic search across the whole database on every query. That is expensive, it is non-deterministic, and when it returns the wrong thing there is no way to see why. Hadano finds the document first, then expands only from what it actually found.
query
|
v
+--------------------------+ stage 1
| keyword search | trigram index + BM25
+--------------------------+ no model, no network
|
| matched chunks
v
+--------------------------+ the only bridge
| seeds at most 3 | between the stages
+------------------------------------------------+
|
v
+--------------------------+ stage 2
| graph walk | one hop, along
+--------------------------+ defined relations
|
v
source passage vectors: designed,
not connected yet
Stage one is keyword search over a character-trigram index ranked by BM25. Trigrams rather than words, so Japanese and other CJK text index correctly without a morphological analyzer sitting in the query path. No model runs, nothing goes over the network, and the same query returns the same result every time.
Stage two is a one-hop walk over the relation graph, seeded by at most three chunks that stage one actually matched. That cap is the whole design. It bounds the cost of a query, and it keeps every result traceable, because there is exactly one bridge between the stages and nothing else crosses it.
Stage two was designed to take vector similarity alongside the graph walk. That half is not connected in this version. It was built against a storage layer that has since been replaced, so today stage two is the graph walk only, and asking for vector expansion returns an error rather than an empty answer that looks like a real one.
02 / The record
Everything, in one table
| Name | Hadano AI Cabinet |
|---|---|
| Category | Local knowledge database, exposed as an MCP server |
| Who it is for | Developers running an AI coding agent that needs to remember things between sessions |
| Interface | Seven MCP tools over stdio. No HTTP API, no SDK, no interface of its own |
| Stage 1 retrieval | Keyword search over a character-trigram index, ranked by BM25. No model, no network, no sampling |
| Stage 2 retrieval | A one-hop walk over the relation graph, seeded by at most three matched chunks |
| Vector search | Designed, not connected in this version. Requesting it returns an error |
| Measured quality | nDCG@10 of 0.58 on SciFact |
| Storage | One file on the machine. No server process, no API key, no GPU |
| Network | None. The local build makes no outbound calls |
| Requirements | Python 3.11 or newer |
| Licence | Apache-2.0 |
| Implementation | Seven of nine specified MCP tools are built and reachable. Two are designed, not built |
| Availability | In development. Not on PyPI. The repository is not public yet |
| Built and operated by | TechJapan LLC, Hadano, Kanagawa, Japan |
| Product site | h-c.ai |
03 / Interface
What an agent can call
| Tool | What it does | Status |
|---|---|---|
| store_document | Store or replace a document; chunks and edges are derived from it | built |
| search_documents | Keyword search over the character-trigram index, ranked by BM25 | built |
| search_knowledge | Expand from seed chunks through the graph, and through vectors if enabled | built |
| query_graph | Walk the relation graph outward from a node | built |
| get_document | Fetch one document by its identifier | built |
| delete_document | Delete a document and everything derived from it | built |
| db_status | Report counts, integrity, embedding mode, and the database path | built |
| compute | Read-only calculations over stored documents, bounded by a row cap and a time budget | designed, not built |
| define_trigger | Register a trigger using built-in functions only | designed, not built |
04 / Why we built it
The same argument we make to clients, applied to ourselves
Everything on the ontology page says that retrieval fails upstream of the model, and that the fix is structure: controlled terminology, defined relations, and a path you can trace. Hadano AI Cabinet is that argument compiled into software — a retrieval engine where the traceable path is the architecture rather than a feature bolted on afterwards.
It is also why the honest accounting on this page matters. A knowledge tool that quietly returns a plausible empty answer is worse than one that errors, because the failure is invisible until a decision has already been made on it.
Read the full technical record at h-c.ai The service version of this