Agentic archiving
An agent that decides where new material belongs in an existing corpus, and puts it there. Capture is cheap and filing is the tax; this is the tax automated. Named here from rocketnotes, whose agent “analyzes snippets from your ‘inbox’ and intelligently files them into the most relevant existing document.”
The maintenance half of the pattern, isolated
llm-wiki‘s claim is that Bush and Engelbart had the design and lacked the labour — nobody could keep the trails current, so the memex stayed a proposal. Agentic archiving is the smallest useful piece of that labour: not writing, not linking, not synthesising, just placement. It is the operation a human does last and worst, on a backlog, badly.
Placed on the corpus’s augment→automate axis it sits between the poles rather than at one:
- llm-wiki-gist — the human directs each edit; the model does the writing.
- agentic archiving — the human captures, the agent decides destination. One decision per snippet, bounded, and visible afterwards in the document tree.
- gbrain — an unattended “dream cycle” daemon rewrites and re-wires the whole corpus between sessions.
What it optimises for, and what that costs
The move only makes sense if destination is the right unit. In a hierarchy it clearly is: a snippet lives in one document, and putting it in the right one is a real decision an agent can make from similarity.
But this lineage’s older claim is that destination is the wrong unit. zettelkasten holds that value is in the links, not the folder — “connection, not a collection” — and associative-trails makes the same point from Bush’s side, that rigid indexing is what fights the mind. An agent that files well is automating the part of the work the method says doesn’t matter, while the part it says does — proposing why this note connects to that one — goes untouched.
Nothing about the mechanism forbids the better version. An agent that reads an inbox snippet and proposes
links with reasons, leaving placement wherever it lands, would be closer to the method and is not
harder to build. No product in this corpus does it, and this hub’s own practice is the counterexample
that shows what it would look like: every ingest here creates [[wikilinks]] and states the connection
in prose, and none of it depends on which directory a page sits in.
Open
- No evaluation, anywhere. Filing accuracy is not reported by rocketnotes and there is no benchmark for it in this corpus.
- What a wrong file costs. A misplaced snippet in a hierarchy is effectively lost until search finds it; a wrong link is visible and reversible. The asymmetry argues for links, and nobody has measured it.
- Whether the agent’s decisions stay reviewable as the corpus grows — the same question gbrain‘s unattended cycle raises at a larger scale.
Related
rocketnotes · llm-wiki · zettelkasten · gbrain · associative-trails · knowledge-graph · synthesis