Troveholm
AI-Assisted
Architecting
Browser companion for durable personal memory
Troveholm remembers the accounts, posts and sources worth returning to, deepens them in the background, and puts what you already know back beside the page you are looking at.
You are on a page you have seen before, or one you have not. Troveholm already knows which. It shows what you have stored about the accounts in front of you, how much of each has been read, and what you decided about them last time — and it does that beside the page, without pulling you out of it.
When something is worth more than a glance, you ask for a deeper read. A worker tab goes and gets it while you carry on browsing, and the result comes back as structured records in your own store rather than as another open tab you will forget.
01 · The product
Returning to a task, not rebuilding it.
Attention across many pages and many sittings is expensive by hand. The same accounts get opened again, the same checks get repeated, and what you concluded last month is somewhere in a tab you closed. The cost is not the reading — it is the reconstruction before the reading can start.
Troveholm holds that context instead. An account you have met is an account it recognizes, with your labels, your notes and your earlier decisions attached. A source you follow is a stored definition, not a bookmark you have to interpret. What you come back to is an updated record with new material already folded into it.
Reading is separated from judging, deliberately and everywhere. Seeing an account does not save it; reading a profile does not approve it. Saving, ignoring, labelling and annotating are four distinct intentions, and the evidence underneath — profile text, post samples, observed relationships — stays a separate layer from what you concluded about it. An observation never quietly becomes a verdict.
02 · How the work gets done
A deeper read, while browsing continues.
A confirmed read becomes a bounded operation with a beginning and an end. Cheetah reserves a worker tab and drives it — navigating, waiting, capturing — so the work happens somewhere other than the page you are reading. A promoted Imprint parser turns each captured page into structured records without calling a model, which means the same page yields the same result at no per-view cost.
What returns is projected into canonical accounts, posts and relationships, each carrying its provenance and its freshness, and each answerable back to the exact parser definition that produced it.
03 · What is built
The machinery underneath.
Server-directed browsing
Troveholm asks for a worker context and issues the collection recipe; reserving the tab, navigating and capturing are handled below the product. It runs no worker pool of its own and stays off tabs you are using.
Model-free extraction
Captured pages are read by parser definitions executed deterministically. No per-view model call, no per-view variability, and no page contents handed to a third party at read time.
Traceable results
Every parser carries its originating project, page type, schema version, promoted artifact version and source commit. Any stored record can be traced back to the definition that produced it.
Owner-scoped memory
Raw captures are preserved and projected into normalized tables keyed to the authenticated owner. A worker receives its owner from the request that created the task and cannot choose one for itself.
Troveholm runs as a local server and a browser extension. Cheetah carries the browser command and worker mechanics beneath it; its parsers come from Imprint definitions promoted through a shared store, each with the provenance chain back to the project that authored it.
Three more of the systems here shaped how it was built without ever being loaded by it. Arboretum carried the architecture review and the multi-document editing behind it. Paladin supplied a deliberately different reading of the same decisions: challenge the assumption, keep the evidence, make the judgment explicit. AWACS kept parallel work legible, with one board showing which session was waiting on a person.