The hippocampal memory prosthesis: a MIMO bypass

The hippocampal transformation

The hippocampus turns experience into durable memory. Along its trisynaptic circuit, the CA3 subfield sends spatiotemporal spike patterns to CA1, which relays a transformed output onward toward neocortex. During memory encoding, this CA3→CA1 transformation is the computation that binds an experience into a storable trace.

If the CA3→CA1 link is damaged — by injury, ischaemia, or disease — encoding fails even when CA3 and CA1 neurons still fire normally. The insight behind the prosthesis is that you do not need to understand what the memory is; you need to reproduce the transformation the tissue can no longer perform.

A nonlinear MIMO model

The MIMO nonlinear memory model — developed over two decades by Berger, Song, Deadwyler, Hampson and colleagues — treats each CA1 output neuron as a nonlinear functional of all recorded CA3 input spike trains. It is fit entirely from paired recordings, with no hand-designed features.

w(t) = k_0 + \sum_{n=1}^{N}\int_0^{M} k_1^{(n)}(\tau)\,x_n(t-\tau)\,d\tau + \sum_{n=1}^{N}\sum_{m=1}^{N}\iint_0^{M} k_2^{(n,m)}(\tau_1,\tau_2)\,x_n(t-\tau_1)\,x_m(t-\tau_2)\,d\tau_1\,d\tau_2

A multi-input Volterra expansion (in practice a sparse generalised Laguerre–Volterra form): the hidden potential w(t) of one output neuron is a first- plus second-order functional of every input spike train x_n; the kernels k capture how inputs and their pairwise interactions drive the output.

y(t) = \begin{cases} 1, & w(t) + \varepsilon(t) \ge \theta \\ 0, & \text{otherwise} \end{cases} \qquad \varepsilon \sim \mathcal{N}(0,\sigma^2)

A probit spiking stage turns the continuous potential into output spikes: the model neuron fires when w(t) plus Gaussian noise crosses threshold θ. In the prosthesis, these predicted spikes become the pattern delivered by stimulation.

From rodent to human

The model was validated in stages. In rodents performing a delayed nonmatch-to-sample task, MIMO-driven stimulation of CA1 restored — and in some conditions enhanced — memory-dependent behaviour after the CA3→CA1 pathway was pharmacologically blocked. The same logic was then tested in nonhuman primates on working-memory tasks.

Under the DARPA Restoring Active Memory program, the approach reached humans: epilepsy patients already implanted with hippocampal depth electrodes for clinical monitoring. Delivering stimulation patterned by each patient's own MIMO model improved short-term and delayed recall relative to no stimulation — again, modest effect sizes, in small, within-subject studies.

Why a bypass, not a recorder

A crucial conceptual point: the prosthesis does not store and replay a memory. It restores the input→output transformation, so the patient's own upstream activity — driven by their own experience — is transformed correctly and handed to intact downstream circuitry. The content still comes from the person; only the broken computation is supplied.

This also clarifies the write target. Stimulating the natural input gateway — for example entorhinal / gateway stimulation — or the CA1 output are different strategies with different demands on the model. Either way, the prosthesis succeeds only if downstream tissue reads its written output as legitimate neural code, delivered via microstimulation with realistic charge and timing.