Stays true.
A self-calibrating in-memory AI accelerator that senses analog drift and restores it in place — so inference accuracy holds for the life of the device.
Mnemostasis is a self-calibrating in-memory AI accelerator. It senses analog conductance drift and restores it in place — so inference accuracy holds across the full life of the device.
Compute-in-memory runs a neural network's matrix math inside the memory itself — weights stored as the conductance of resistive cells, multiply-accumulate done in the physics. The speed and energy wins are real. So are three failure modes that quietly erode them.
Cell conductance shifts over time and under read stress. A weight set to G becomes G ± ΔG, and accuracy slides downhill with no built-in correction.
Resistance along long rows and columns skews the summed current — error grows with array size, capping how large a useful tile can be.
Analog-to-digital readout tends to dominate energy and area, eroding the very efficiency that made in-memory compute worth building.
Without compensation, top-1 accuracy slides as cells drift over billions of inferences. Mnemostasis holds it. Drag across the device lifetime to compare.
Curves are illustrative of the compensation behaviour, not measured silicon results.
| Digital accelerator | Analog CIM (uncompensated) | Mnemostasis | |
|---|---|---|---|
| Energy / inference | High | Low | Low |
| Accuracy over life | Stable | Degrades with drift | ● Held in tolerance |
| Array scale | Power-bound | IR-drop-bound | Segmented bitlines |
| Self-calibration | — | None | ● In-situ, per tile |
Each tile carries its own reference cells of known conductance. Mnemostasis reads them, measures how far the array has drifted, and restores it in place — during idle cycles, while neighbouring tiles keep inferring.
Known-conductance reference cells share the tile's process corner and temperature, so their deviation is a true reading of the drift the weights are feeling.
When the drift metric crosses threshold, partial programming pulses with read-back-and-verify walk each cell's conductance toward target — no off-array weight reload for routine drift.
A controller schedules restoration by drift metric or inference count, within a pulse budget that never starves inference — so accuracy stays inside a bounded tolerance for the device's life.
Local, in-situ drift sensing instead of a global guess. Every tile measures its own drift against cells it can trust — they live in the same silicon.
Charge-domain accumulators break each column into segments, bounding IR drop and letting one lower-resolution converter be shared.
Maps layers to tiles and times restoration by drift metric or accumulated inference count, catching slow drift before accuracy crosses the floor.
Representative figures for a reference design; final parameters depend on device and process selection.
Strip away the silicon and Mnemostasis is a feedback loop of the oldest kind: it measures how far a system has drifted from its setpoint and acts to close the gap — the same negative feedback that holds a body’s temperature steady. Norbert Wiener named the principle in 1948 — cybernetics, the study of control and communication in the animal and the machine.
Sense the error, drive it to zero. The reference column measures drift; restoration cancels it. The loop’s whole job is to keep the error small.
A variable held steady against disturbance. Here the variable is conductance, the disturbance is drift, and the setpoint is the programmed weight.
Cross a threshold, trip a corrective step. Mnemostasis holds accuracy inside tolerance in exactly that way.
Lineage, not label: engineers will file this under control systems and neuromorphic design — but the governing idea is pure cybernetics, a machine that regulates itself toward equilibrium.
Drift compensation earns its keep wherever a device runs for years and can't be recalled to reload its weights.
A decade in the field across wide thermal cycling. Perception accuracy can't be allowed to quietly drift between service intervals.
Devices that can't be retrieved to recalibrate. Inference has to stay trustworthy for the life of the implant.
24/7 inference at the edge means drift accumulates fast. Self-restoration keeps lines and sensors honest without downtime.
Harsh environments, long deployments, zero maintenance windows — exactly where uncompensated analog accuracy fails.
Provisional patent in preparation; the protectable combination defined and the §101 posture set.
Behavioural and device-level models of drift and restoration, validating the accuracy-retention claim.
A small-array silicon prototype proving the sense-and-restore loop in real cells.
Full-tile integration with segmented bitlines and the shared converter.
Qualification and integration into partner accelerators and edge platforms.
The defensible position isn't any single part — it's the combination: an on-tile reference column, in-situ partial-pulse restoration with write-verify, and inference-count-aware scheduling, working together inside the array.
Mnemostasis is framed for patent eligibility as a hardware improvement — a technological solution to a technological problem — rather than as a claim over the underlying neural-network math. Apparatus claims lead; the restoration loop, not the matrix multiply, carries the inventive weight.
Patent-pending design · provisional application in preparation. Subject to prior-art and freedom-to-operate review.
No. Restoration runs per tile during that tile's idle intervals, within a pulse budget, while neighbouring tiles keep inferring. The scheduler is built so calibration never starves the workload.
It's technology-agnostic. The reference-column and partial-pulse approach applies to RRAM, phase-change memory, or floating-gate cells — anything where a weight is held as an analog conductance that can drift and be re-tuned.
Inference. Weights are loaded once; Mnemostasis keeps them true. The invention deliberately does not claim the training math — that framing also keeps it on the right side of patent-eligibility rules for hardware.
Reference cells of known conductance are read continuously; their deviation gives a true per-tile drift metric. When it crosses threshold, restoration pulls cells back to target. Accuracy is held inside a stated tolerance rather than allowed to decay.
A combination claim led by apparatus claims: reference-column sensing plus in-situ write-verify restoration plus inference-count-aware scheduling. The inventive weight sits on the restoration loop — a concrete hardware improvement — not on the matrix multiply.
Get in touch
For partnership, licensing, or technical detail on the architecture and its IP posture — we'd be glad to talk.
contact@mnemostasis.com