A model carries what it learned across a turn or a session by compressing it, and the usual instinct is to keep the answer and drop the work. This is about what that costs. A memory that keeps a conclusion but sheds the source it came from cannot be corrected, because there is nothing left to recompute from. The error survives every attempt to fix it. We call it brittle memory.
In an agent the same failure gets sharper: an agent that compresses away its own actions cannot act on its own past. Below you can watch the mechanism, then see it measured on real frontier models.
Two agents play the same board, with the same events, under the same memory budget. The only difference is what their compression keeps: the recomputable source, or the salient conclusion. Press play, and switch the scenario to see the boundary of the fix.
A deterministic illustration of the mechanism, with no model calls.
The demo is built to teach the shape. To check it is real, the same setup runs on live models commanding a game of Battleship from a compressed memory: every turn is a fresh context carrying only the memory of the game so far. With source-first compression the coordinate record survives, so the model can avoid firing the same cell twice. With lossy compression it is gone. We ran three models, eight games each: a pilot, not a survey.
| Model | re-fire, source-first | re-fire, lossy | progress, lossy |
|---|---|---|---|
| Opus 4.8 | 0% | 91% | 0.4 hits/game |
| Sonnet 4.6 | 0% | 77% | 1.5 hits/game |
| Llama-3.1-8B | 29% | 42% | 2.6 hits/game |
Re-fire rate: how often the model fired a cell it had already fired. Eight games per cell, judge-free scoring, the lossy and source-first memories matched to the same byte budget.
The rule, in one line: preserve what cannot be recomputed, and mark whether that preserved source is complete.
Paper in preparation. github.com/collapseindex · collapseindex.org