Are you thinking of Serial Box?

- mostly-externally-sourced stable identifiers for objects involved in those events -- so ISBNs for books, commit hashes for code changes, etc. Keep the details about those objects in external DBs and reference them as needed. (Some objects would be internal to this store, though -- my notes, etc.)

- a store of my intents -- I want to learn X, I want to run a 10k -- tracked over time, to help interpret the event stream. These should be at least explicitly specifiable, but they should also be inferable from activity (with oversight from the person)

- event sourcing, capturing relatively unstructured descriptions of events as they happen and making them available for later processing (and re-processing as use cases change).

This is a hard question, because the shape of a data store _should_ vary based on how you want to use it -- and a data lake for a life could serve a wide variety of both known-upfront and unforeseen uses. Ignoring a whole raft of tradeoffs, I'd want to start from:

I love it as well! Though I was surprised by how much less I liked the back half of AUs2 after what seems like it should be a minor change to the flow of the show (vague to avoid spoilers)

Also, I'm strongly reminded of Lindley Darden's work on mechanisms in science (especially biology); it's been a very long time since I was in her class in grad school, so it's probably time for some reading to refresh my memory.

So, yes to being able to expose the full context to a model, but please give me controls so I can choose _not_ to, and have it know about the subset that it needs to do the job I want it to do most efficiently.

(There's already an agent whose job it is to manage _all_ my context: me.)

context of my life to each group would be counter-productive to the unifying thread of each group.

I think it's the same with AI models -- we'll use different instances of models for different things, and we'll want those models to have access to different subsets of our overall context.

both intentionally and accidentally, and for the most part that's good. I don't want my sport friends to be distracted by the most recent book I've read while we're playing a game.

There's absolutely a place for crossing those boundaries occasionally, but offering the full

Nice provocation! Taking it a little further, I think the benefit is being _able_ to offer models any specific subset of our context, not _always_ offering _all_ of it.

Take groups of friends: I have my job friends and my book friends and my sport friends. I'm a different person to each group,

This reminds me (tangentially) of McNamara's work on what resources work better for people with differing levels of experience in a domain -- clearer texts help novices, but are worse for people with more background knowledge (e.g., doi.org/10.1207/s153...)

You are in luck! store.dropout.tv/products/inv...

+1, lemme know if you want any help!