Most agent memory is a key-value store with embeddings: what you put in is what you get back, forever, at the same rank. Graphonomous is a loop instead. An action reports its outcome, the confidence on the nodes that informed it moves, the topology of that neighbourhood changes, and the next retrieval ranks differently — with no model retrained and no weights touched. Whether that actually makes an agent better is the open question. What is settled is that the mechanism exists, runs, and can be read.
The registry has it, the binary downloads, and the server answers on stdio. It is also the honest version: the installer advertises four targets and only one of them has ever been built. If you are on a Mac, this will fail, and it will fail during npm i rather than later.
# 1. the published package — linux-x64 only, see the table below
npm i -g graphonomous
# 2. ~/.mcp.json , or your client's MCP settings
{
"mcpServers": {
"graphonomous": { "command": "npx", "args": ["-y","graphonomous","--db","~/.graphonomous/knowledge.db"] }
}
}
# 3. restart the client. Node.js 18 or newer; the graph is a
# SQLite file at ~/.graphonomous/knowledge.db and does not leave the machine.Requested with curl against https://github.com/c-u-l8er/graphonomous/releases/download/v0.4.2/ on 2026-08-16, one request per target, asset name graphonomous-v0.4.2-<target>.tar.gz. The status column is the HTTP code that came back.
| Target | HTTP | Result of npm i -g graphonomous |
|---|---|---|
| linux-x64 | 200 | installs |
| linux-arm64 | 404 | no asset — install exits 1 |
| darwin-x64 | 404 | no asset — install exits 1 |
| darwin-arm64 | 404 | no asset — install exits 1 |
The gap is not a mystery and it is not hard: npm/scripts/resolve-platform.js maps two platforms by two architectures, and the release workflow has only ever produced one of the four archives. Until the other three exist, the reachable audience for this package is Linux on x86-64. That is the single largest thing standing between this project and someone outside it trying it.
The engine exposes one tool per phase of the loop rather than one tool per capability. Each machine dispatches on an action parameter, so a client sees a short menu of verbs and the long menu lives underneath. The phases are the same five PULSE names every loop in this portfolio declares.
Every action above is read out of the @valid_actions list in its machine module at build time. When a machine gains one, this page gains one; when this page and the modules disagree, the build stops. That is how the previously published table came to be short by two — retrieve/replay and act/store_trace both existed in the code and neither was on the page.
Those two halves belong in one sentence, and separating them is how a page starts lying. The engine's own suite passes. The engine's own suite runs against the tree in this repository, and the tree is a version ahead of anything the registry serves.
A gate is either approved with its evidence, its reviewer and its date, or it is pending. There is no third state, and the publication gate refuses an artifact whose ledger claims approval without all three fields.
records/witness.json#tests_total. Reviewer: mix test (Elixir 1.19.4 / OTP 28). Date: 2026-08-16.records/witness.json#install_targets. Reviewer: HTTP probe of github.com/c-u-l8er/graphonomous/releases/download/v0.4.2/. Date: 2026-08-16.The revision this replaces was 215 KB and stated most of its figures in the present tense with nothing behind them. These came off:
455 tests — The suite reports 577. 455 was the count at some earlier commit and had been sitting in the present tense.29 actions — The 5 machine modules declare 31 valid actions between them. The published table also omitted retrieve/replay and act/store_trace, so it was internally short by the same two.29 individual tools — Same stale denominator, in the sentence that argues why the machines exist.29 MCP tools — Same stale denominator, attached to a 100% pass-rate claim.macOS or Linux — Published as an install requirement. Only linux-x64 has a release asset; the other advertised targets all return 404 and the postinstall exits non-zero. The largest single audience for this page could not install it, and the page said they could.No other agent memory system does this — A claim about every other system in a field, with no survey behind it. Nothing on this site establishes what other systems do.96.6 — A retrieval-accuracy figure attributed to a third-party system with no citation, in a comparison table this site cannot witness.Only the first four were wrong at the time of writing; the rest were removed for want of a witness, which is a different and smaller charge. The install requirement is the one that mattered. A reader on a Mac was told the package supported their machine, typed the command, and watched the postinstall exit non-zero — and the page that sent them there was the project's best acquisition surface. A link that returns 200 can still be dead, and so can a package that resolves.
The fix is structural rather than careful. This page is generated: every figure on it is emitted from records/witness.json, each entry names the command that produced it, and launch-gate.mjs refuses to publish an artifact that reinstates any string above outside this paragraph. A number can no longer be typed onto this site by hand.
A page may only ask you to do what its evidence has earned. There is a deployed artifact, so this page may ask you to use it — on the one platform it exists for. The version being described is only in the tree, so about that, it may ask you to read and to run the tests, and nothing more.
lib/graphonomous/mcp/machines/. Each declares its own @valid_actions, and that list is what this page prints.Run the testsClone, mix deps.get, mix test. Needs Elixir ~> 1.17. If you get a number other than 577, that is the most useful thing anyone could send us.This page publishes an install table in which three of the four advertised platforms return 404, and a ledger in which nobody has yet installed the published package on a clean machine and spoken MCP to the result. A reader telling us the linux-x64 tarball did, or did not, work for them is the shortest path this domain has to the external rung — the one rung nobody here can award themselves. A result, a stack trace, or a figure on this page you think is wrong: same channel, all three welcome.
It is a real <form> with an action and method="POST", so it submits with JavaScript disabled; /contact.js only upgrades it to an inline reply, and that reply prints sent only when the endpoint actually answers 2xx. A form that says thank-you on submit and drops the message is the exact failure this site is about. There is still no email address here, deliberately. Anything better filed as a bug goes to the issue tracker.