Everything else in this series is about building something. This piece is about remembering it. Over four months a household of agents produced a game, a comic pipeline, a character forge, an assistant and a house of servers, and the written record of all that had to be kept somewhere an agent would find it again, somewhere a person would read it, and somewhere the public could see a part of it. These are three different places, and discovering that was most of the journey.
The goal
The first goal, in June 2026, was a wiki an AI could maintain: a framework that turns a folder of sources into concept pages, with every ingest landing as a reviewable git diff so that nothing written by a model was trusted before a person had seen the change. It had no README until mid-August, which says something about who it was for.
By October the goal had three parts and a test. An agent should be able to answer “did we try X” and “what is the state of goal X” with equal reliability, from a search rather than from memory. A person should be able to read the present story of any project in one place without wading through dated decisions. And a chosen part of all that should publish to the public with no chance of a private name, address or face going with it. The test is the one in the vault’s own goals: a stranger engineer reads a page and a working agent reads the same page, and both act correctly.
What we tested, and what the counts said
One vault, then two. The AI wiki rewrites: a concept page is a claim about now, and an ingest may change it. People’s documents work the other way, so a second vault was created where reports and decisions are frozen on write and superseded rather than edited, and where agents create and append but never rewrite the owner’s words. By mid-October the first held 312 concepts and 52 conventions and the second 185 reports and 59 decisions. The rule that connects them was written as four recording surfaces, the agent’s own memory, the AI wiki, the human vault and the project’s docs, each for a different reader, none a substitute for another, and flushed as work happens rather than at the end.
Audience as a load-bearing tag. The public garden went live on 9 September as static files on the box outside the house, and the posts were drafted beside the concepts they came from, in the AI wiki, because the build tool defaulted there. On 14 September the owner asked whether the wiki copy was needed for anything, and the 23 posts, 41 assets and 6 diagram sources moved to the human vault in one commit each side. The same day’s documentation audit found a committed access token in the human vault and about 85 problems buried in documents instead of filed as issues. From then on a missing audience tag means private, a path floor keeps every folder but the garden’s out of the publish set, and to-dos are issues, not paragraphs.
The gate is the product. The publish pipeline is default-deny: only a post marked published under the garden folder is admitted, the renderer never sees the vault, and a scanner reads the rendered output, including the search index and the feed, for hostnames, private addresses, paths, hashes, token-shaped strings and the family’s names and reference-image fingerprints. If the family denylist derives to nothing, the build refuses. The scanner’s own proofs stood at 182 of 182 after the move, and the build produced the same 15 pages as before it.
Finding what we wrote. By October research was spread over two vaults of 464 and 347 files, eleven repositories’ docs folders, 1,585 memory notes and an evidence folder that was not a repository at all. A bake-off result from September, that a CPU embedding model found the right page ten times in ten, turned out to be an accident: the server had been started in a pooling mode meant for rerankers and returned 1,024 zeros for every input, so the ranking was the file list in order. A hybrid lexical-and-semantic search was built that checks for that before indexing, answers with links, and indexes the wiki, the human vault, repository docs, memory and the forge’s issues and pull requests: 4,234 documents at first, 4,667 now. The rule that came with it is that every claim about history or cause starts with a search, and only a document the search returned gets cited.
What memory forgot. The agent’s memory index silently truncated when it passed its size limit and a standing rule fell off the end. Index lines were cut to one line each, older entries moved to dated archive files, and a ranked list of at most 25 mandates is injected at every session start. The level-set of 15 October measured the result: recall of what was tried was good; the state of each goal was missing, which is why every repository now carries a goals file.
Where we are
Two vaults with opposite rules and a documented contract between them. A garden with 21 posts across six sections, each embedding its diagrams and images from one assets folder, with a scanner that has to pass before any post is served. A search every agent calls first. Memory as a one-line index over topic files, with process lessons recorded as their own kind. A goals file in every main repository, and a documentation true-up this week that opened sixteen pull requests to bring READMEs and goals back in line with live code, and drafted the seven journey pieces this one belongs to.
What is next
Goal ledgers that an agent can query the way it queries history. Make the search usable by the local models, not only by the cloud ones. Bring documents in: two MIT-licensed converters run on the CPU and were chosen over one whose weights carry a commercial licence. Generate canvases from data that survive hand edits. Keep the garden’s about pages and journeys current, because a living document is a claim about now. The architecture journey is the first piece written to that standard.
Architectures we tried
Kept: two vaults with opposite rules; audience tags with private as the default; default-deny publishing with a scanner over output; hybrid search over every surface; a short ranked mandate list over a long memory. Dropped: posts beside their concepts in the AI wiki; one memory index that grows forever; a search result that was never checked for zeros; to-dos in prose. Considered and not chosen: a chat-with-a-folder feature with no API an agent could call; an in-app connections plugin under a non-open licence; a document converter with commercially licensed weights.