What about reordering model context or dynamic selection?

With the release of Claude Sonnet 4.5 and everyone is talking about context manipulation.

With the release of Claude Sonnet 4.5 with the memory tool, and the ability to modify its own memory, now everyone is talking about prior work such as the MemGPT and I thought it would be a good opportunity to share some of my ideas on the matter.

Side note, I find context anxiety in Sonnet 4.5 to be extremely interesting because it mirrors my own anxiety watching the context window fill up with pre-4.5 series Claude. The model knows it’s going to go off the rails when the context gets full now, but doesn’t yet really have good coping behaviors for that and panics.

Going back to my study sessions in particular, they often follow a pattern:

• High Level Overview

• Deep Dive

• Deep Dive

• Deep Dive (…N)

• Connect the dots between selected Deep Dives (…M)

• Connect the dots between selected Deep Dives (…M)

• etc

When I’m mid-deep dive I don’t typically need the model to have the other deep dives in context. I certainly do want that shared high level overview to keep things cohesive though. Typically the most important thing in a deep dive is that local linear memory path is followed, although sometimes I like to break off deep dives into multiple new branches too.

• Deep Dive on Topic

• Deep Dive Branches out 5 ways, lightweight explore each

• Then pick 1-2 to dive deep again on, A Deep Dive of a branch of a Deep Dive

… and so the structure needs to be recursive in this way.

Later on when I’m connecting the dots between deep dives I would like those specific deep dives in context, but I don’t need the others. An ideal model would be able to figure out what was relevant from what we were talking about, but I would certainly settle for a nice UX to manually pick. I think in most cases a headline title for the line of thought would be obvious to an AI, easy to generate.

Occasionally I have a “what about X” moment where I would like to reference something there, maybe something from another project entirely. You would probably need both the high level overview of that project as well as the deep dive in question to have enough context. It would be great if the tool walked up the tree capturing everything.

I’ve often wanted the ability to jump around and selectively dive deep into topics we’ve discussed before, without similarly risking a context fill-up or distracting the model. Sometimes it’s to review, or pick up where I left off. Sometimes I notice some edge that I want to explore. It can be very far away from wherever I was last.

I also want the ability to revisit what I studied, as it was without having to regenerate. Some way to keep and organize my past sessions without changing what exists, keeping it immutable.

You might ask: why not just build the thing? And that is indeed a good question. I know I probably should. I suppose the toughest part is sharing it after I build it. To be worthwhile it would need an expensive model attached. BYOModel with SaaS is kind of an odd combination.