Activity Frames Fix Agent Memory by Making It Deterministic
A new paper introduces deterministic screen-activity representations that let agents remember and replay UI interactions exactly, solving the memory bloat that plagues current agentic systems.
Why it earned a slot
August 12, 2026. The field surfaced two papers that tell opposite sides of the same story. One is the Catastrophic Remembering paper from Kushal Chakrabarti: agentic READMEs like CLAUDE.md grow without bound, tripling over their lifetime (+226%) with +4.9 net instructions per commit. The other is Activity Frames, which proposes a deterministic screen-activity representation that lets agents remember and replay UI interactions like a video-game save-state. I think the field is wrong to treat agent memory as a language problem. It's a state problem. Every time you append a new instruction to a prompt, you're adding a constraint that might conflict with the last one. The Catastrophic Remembering paper shows exactly this: the deletion cost is O(2^|D|) in a prompt of |D| instructions, so no one deletes anything. The file grows until someone rewrites it wholesale. Activity Frames cuts the knot. Instead of accumulating textual instructions, you compile the screen state and the action taken into a deterministic record. You can replay the exact sequence of screens and actions. You can branch and roll back without accumulating garbage. The paper claims this representation is deterministic — identical input leads to identical output. That's crucial for debugging. I don't know how Activity Frames handles dynamic content like pop-ups that change the screen layout mid-action. The paper's abstract doesn't address that edge case, and I'd need to see the full results on their benchmarks. But the core idea is sound. Right now, debugging an agent that has a 226%-grown CLAUDE.md is like debugging a distributed system with no logs. You don't know which instruction caused the failure. Activity Frames gives you a replay log. The Catastrophic Remembering paper shows that prompt comments can remove 99.3% of excess instructions. That's a fix, but it's still a textual fix. Activity Frames proposes a structural fix: change the representation, not the content. That's a stronger approach. This is the kind of engineering work that matters more than another model scaling trick. If we want agents that can run for months without accumulating poisonous instructions, we need deterministic memory, not bigger prompts. The paper is worth reading, and I hope the code is released soon.