A Memory That Only Looks Like One: A Hybrid Vision-Language Model Relays the Picture Rather Than Storing It

A hybrid vision-language model carries two kinds of memory. Its attention layers keep a growing cache: every token the model reads, including every image patch, is stored as a key-value entry that later tokens can look up. Its recurrent layers keep a single state of fixed size, and every new token overwrites part of that state. Prior work proved that a fixed-size state must eventually lose information, but not when a given model forgets, what it keeps while forgetting, or what breaks downstream. We answer these three questions for Zamba2-VL-7B, whose 81 decoder layers all carry a Mamba2 mixer while 13 also carry attention. First, just after reading an image the state holds the gist, not the details: a linear probe reads which object was present at $92.5\%$ on the best layer against $12.5\%$ chance, but can barely read how many objects there were or where they sat. Second, how fast the state forgets can be computed from the weights alone: each head has a half-life of $\ln 2/(|A|\bar{\Delta })$ tokens, with medians of $6.6$, $2.9$ and $2.6$ at three scales, nothing fitted. Third, and centrally, the memory a probe appears to find at distance is mostly not memory. With write amplitude held fixed, heads with a $2.3$-token half-life still decode identity at $26.6\%$ sixty-four tokens after the image, where their own parameters retain $4\times 10^{-9}$ of what was written. Evicting the image's key-value entries, without touching the recurrent state, drops the fast tercile from $51.6\%$ to $7.8\%$ at thirty-two tokens, below chance, while evicting an equal amount of pre-image text moves the readout the other way; at matched amplitude, eviction costs fast heads $2.4$ times what it costs slow ones. Causally, exchanging all 81 recurrent states between runs that saw different images leaves $100$ of $100$ answers following the host's own image; exchanging seven of thirteen attention caches moves $94$ to the donor's. The recurrent channel relays what attention holds; it does not store the picture. The redundancy premise behind visual token pruning survives down to a quarter of the entries, but the cost of pruning harder roughly doubles with distance from the image.