A hippocampal · memristive · neuromorphic architecture where STDP, replay-driven consolidation, two-phase processing and neuromodulation are not programmed but emerge from the intrinsic dynamics of a physical material.
· the first system where adaptation is a material property ·
MEMPHIS will establish proof-of-principle that hippocampal consolidation mechanisms — STDP, replay-driven memory consolidation, two-phase processing, and neuromodulation — can emerge from the intrinsic dynamics of a self-organising memristive substrate. A neuromorphic chip in which adaptation is a material property, not an algorithmic feature layered over silicon.
The decisive experiment is a CA3 ↔ CA1 module that improves task performance after an offline phase — without further training data, with energy per synaptic operation at least two orders of magnitude below GPU baselines and below the ~10 pJ/event of Loihi and TrueNorth. The work targets TRL 4 in 36 months, with three objectives, a five-partner consortium, and a route through EIC Transition to industrial pilots in edge AI, autonomous robotics and implantable neurotechnology.
Sister microsite to SYMPHONY within the Synaptic Cartography series — both bet that the next computing era is biological, structured and auditable, not larger and more centralised.
The long-term vision is a new class of artificial intelligence: spiking, adaptive, and physically embodied computation inspired by selected functional principles of the mammalian hippocampus. Not biological intelligence in full — a more specific and credible objective. Demonstrate that biologically grounded mechanisms can be realised as material properties of a neuromorphic substrate.
Adaptation as material property, not algorithmic feature. STDP, replay and two-phase consolidation emerge from the substrate's physical dynamics. Inherently compatible with multimodal sensory input and continuous operation.
Consolidation and energy efficiency arise from the same dynamics rather than from separate optimisations — a single architectural property, inseparable in principle.
Systems that gain task-relevant knowledge over their operational lifetime without recharging or retraining. Not achievable by combining existing efficient hardware with existing adaptive algorithms; the two properties have to be architecturally fused.
Two fields independently reached a point of convergence and stopped. Bio-inspired approaches demonstrated real advantages in continual learning, but were implemented on deterministic digital hardware that preserved the von-Neumann separation of memory and computation. Memristive devices reached synaptic plasticity characteristics in the lab, but were used in isolation — without the network-level mechanisms (replay, neuromodulation, two-phase processing) that make biological learning work.
Bringing the two streams together is not a software port. It is a physical co-design problem. The substrate has to be the algorithm.
STDP, replay-driven consolidation, two-phase processing and neuromodulation are interdependent components of a single biological memory mechanism — yet every existing approach implemented each in isolation. MEMPHIS is the first system where all four emerge from the intrinsic dynamics of a single physical substrate.
Prior work treated memristive stochasticity, variability and nonlinear conductance updates as non-idealities to be mitigated. MEMPHIS inverts the premise: those device-level properties become the substrate of biological-like learning at network scale, rather than something to fight.
Three classical barriers — STDP voltage-time integration vs. spike timing, CA3 attractor connectivity statistics, and memristive switching scales — are addressed not by waveform engineering but by physically co-designing devices and circuit motifs so that the hippocampal computational primitives are intrinsic to the material.
A computational model reproducing associative encoding, novelty-gated updating, bidirectional replay, two-phase dynamics and neuromodulatory control. Verified against in vivo rat-hippocampus recordings.
Threshold · ≥ 80 % of neurons match in vivo electrophysiology (firing rates, STDP weight changes, sharp-wave ripple statistics)
Memristive devices with synaptic properties physically compatible with the CA3 ↔ CA1 model — STDP timescales matching biology, physically differentiated excitatory and inhibitory analogues, stochastic network topology.
Threshold · Switching energy < 10 fJ per synaptic event · stable, reproducible switching · verified learning curves · power-law noise dynamics
Demonstrate that hippocampal consolidation mechanisms emerge from the intrinsic dynamics of the memristive CA3 ↔ CA1 module. The decisive experiment.
Threshold · Measurable task-performance improvement after offline phase vs. online-only baseline · energy ≥ 2 orders below GPU · competitive with existing neuromorphic chips · validated on a robotic navigation task
A small-scale memristive spiking network (the CA3 ↔ CA1 module visible at the centre of the chip plate) performs associative recall on a held-out test set. The system then enters an offline phase — no external input, only intrinsic dynamics replaying the stored traces. Memristive thresholds shift, redundant weights fade, salient patterns are reinforced. After the offline phase, the same test set is re-evaluated. The expected result: improved recall performance achieved without further training data, with energy per synaptic operation at least two orders of magnitude below the GPU baseline.
This is the proof of principle. It demonstrates — in physical hardware, not in simulation — that adaptive learning and memory optimisation can emerge from intrinsic system dynamics. If the experiment lands, the architectural principle generalises; if it does not, the failure mode is informative for the next-generation memristive design.
The target switching energy — below 10 fJ per 100×100 nm device — sits three orders of magnitude below existing neuromorphic platforms (Intel Loihi, IBM TrueNorth), six orders below GPU-based AI, and approaches the biological benchmark of ~100 fJ per synaptic event. The corollary is that entire product categories become physically possible that today are not.
Battery-powered or energy-harvesting devices that need on-device adaptation — environmental sensors, agricultural monitoring, condition-based maintenance. Today's silicon either does inference (small, dumb) or learning (large, plugged-in). MEMPHIS does both inside the energy envelope of a coin cell.
Mobile platforms — drones, agricultural robots, logistics, flexible manufacturing — that cannot afford cloud retraining and cannot afford catastrophic forgetting. The hippocampal-replay primitive maps onto exactly this constraint. WP2 includes a robotic-navigation validation task.
Closed-loop neurotechnology — sense, interpret, act, learn — running for years on a milliwatt budget without cloud retraining or replacement surgery. Privacy-preserving by construction. MEMPHIS targets the architectural precursor; full clinical pathway is downstream.
Validated memristive design primitives establish a European foundation for next-generation neuromorphic hardware — reducing dependence on imported and energy-intensive solutions. High-value EU jobs in hardware design, AI engineering, and advanced robotics.
Whether memristive devices can be matched and stabilised at the precision required by the replay-driven dynamics. The biology demands a tighter device-to-device tolerance than today’s memristive arrays routinely deliver. The engineering question is whether self-organisation can close that gap inside the operating regime, not whether it must. MEMPHIS’s second breakthrough — treating stochasticity as substrate rather than noise — is the bet that it can. Proof of principle for the CA3 ↔ CA1 module is what tests the bet.
10 fJ/event vs. 10 pJ/event for Loihi / TrueNorth · vs. ~1 nJ/event GPU-equivalent for foundation-model inference. Three orders of magnitude below state-of-the-art neuromorphic, six below GPU. The CFO maths writes itself for any edge or implantable deployment.
MEMPHIS does not need a new memristive material to win — it needs the architectural property that bio-replay imposes on the device array. The device roadmap is being pushed independently by half a dozen industrial labs. MEMPHIS rides that wave with a 3–5 year lead in the system-level integration.
MEMPHIS is the hardware substrate that SYMPHONY's neuromodulated software substrate eventually needs. Co-developed roadmaps, shared editorial discipline, one consortium philosophy. A portfolio bet across two pillars of the post-von-Neumann era, not a single shot.
MEMPHIS is at TRL 1–4. The deliverable is a proof of principle, not a product. The value is in the IP position, the device-co-design relationship with the memristive labs that are 3–5 years ahead of where the field expects them to be, and the architectural priority over any team that arrives at the insight later.