The Four Dials Every AI-and-Archive Build Answers, Whether You Write It Down or Not
Every system that pairs a person with a machine intelligence and a growing archive makes four decisions. Not eventually — immediately, on day one, usually without anyone choosing them on purpose. Skip the choosing and the system chooses for you, by default, and the defaults are almost never the right setting for the specific person using them.
We call them the Four Dials: review cadence, summarization autonomy, taxonomy style, and memory-editing rights. Name them on purpose and a surprising amount of what makes an AI-and-archive system curdle after six months simply stops happening. Leave them unnamed and you've still set them — you just don't know what you set, or why it's wrong for you specifically.
Review cadence
How much of what the system captures does a human actually read before it's treated as true?
- Gate everything. Nothing is real until you've read it. Slow, safe, right for anything meant to be a faithful record of you in your own words.
- Spot-check. Sample periodically, trust the average case, catch drift before it compounds.
- Spec-gated. Bounded by a written rule instead of a per-item approval — the system polices itself against a standard you set once, not a standard you re-apply every time.
The right setting isn't a preference, it's a readout of what's at stake. A technical reference archive can run loose. Anything that's supposed to be you, in your own words, usually can't.
Summarization autonomy
Does the model condense material under supervision, or is it allowed to distill unsupervised and hand you the result?
Instrumental material — meeting notes, a contact list, a project's status log — tolerates the second just fine; nobody's identity is riding on whether the summary is exactly right. Anything identity-bearing usually doesn't, for a specific reason worth naming: an unsupervised summary of you is a claim about you, made without your sign-off, and claims like that compound. Ten quietly-wrong summaries in, the archive is describing someone who resembles you less than it used to, and nobody decided that on purpose either.
Taxonomy style
Three real options, not two:
- Fixed categories, decided up front. Clean, predictable, and wrong the moment life produces something that doesn't fit a box you built before you knew it was coming.
- Flat and associative — new material links to whatever it's nearest to, no imposed hierarchy. This is closer to how a mind that's actually working feels from the inside.
- Actionability-based — Projects, Areas, Resources, Archives, sorted by what you're going to do with something rather than what it is.
The tell is in what already died. If someone's abandoned systems all failed from rigid folders imposed too early, handing them fixed categories again — because fixed categories are what you personally know how to build — isn't a fresh start, it's the same funeral with a new coffin.
Memory-editing rights
Does the AI get a self-editing memory layer at all, and if it does, what actually bounds it? Not "a review queue for every entry" — a written specification: a type taxonomy, an explicit exclusion list, a clear account of what counts as memory-worthy and what doesn't. The loosest setting on this dial isn't "no rule." It's "no per-entry approval, but a tight rule for what's allowed to become a memory in the first place." Those are very different things, and confusing them is where this dial usually goes wrong.
MONET runs all four at maximum friction — on purpose, and not as a template
The system this framework was extracted from gates every entry, distills only under supervision, holds a fixed taxonomy, and grants no unbounded self-editing. That's the correct setting for a corpus that's a faithful, ongoing record of one specific person's whole reflective life. It is not the correct setting to copy onto a client's research notes, a team's shared knowledge base, or anyone whose archive isn't that kind of record. The framework exists precisely so the four questions get asked fresh, every time, instead of inherited from whatever system happened to be lying around.
Where this sits among real precedent
None of this was invented from nothing, and it's worth naming what it's actually in conversation with. Andrej Karpathy's LLM Wiki pattern — an LLM reads sources once and distills them into a compounding wiki, cross-references already built rather than inferred fresh each session — sits at the loose end of both the summarization and review dials: fast, cheap, built for material where being mostly right, fast, beats being exactly right, slow. Tiago Forte's PARA/CODE gives the actionability-based taxonomy its clearest published form. Niklas Luhmann's Zettelkasten is the flat-associative option in its oldest, most tested shape — a card index that out-thought its own owner by making connections he hadn't consciously drawn. MemGPT and its successor, Letta, are the memory-editing dial made explicit as an actual research question: what does it mean to let a model manage its own paged, persistent memory, and what has to bound that so it doesn't drift.
Each of those is a considered, defensible setting. None of them is a universal default, including MONET's own. The work isn't picking the "best" answer to the four questions. It's noticing that they're being asked at all, and answering them for the specific person and the specific archive in front of you — which is the part that doesn't come pre-built in any of the four traditions above, and the part an unexamined system will always get wrong first.
Full worksheet available as part of the Extended Mind Buildout — see the offers or apply directly.
← Back to aiBLOG