Priver Mneme · On-device memory eyewear

Glasses that remember for you. They can't tell.

Everyday glasses that quietly build a private memory of your days — surfacing the name, the place, the promise the instant you look. The raw camera and audio never leave the lens. No cloud. No retraining. Only recall.

<100ms
on-device recall
0 bytes
raw sensor data sent
100%
memory you can edit & erase
MAYA · entity met 14 days ago · 0.91 “the Kyoto trip in spring”
The premise

Today's smart glasses send what you see to someone else's computer. Mneme keeps it inside the frame — and gives the keys to you.

Persistent, personal memory has been impossible to put in eyewear without breaking one of three constraints: latency, battery, or privacy. Cloud round-trips are slow and leak your life. On-device generative models drain the battery and overheat. Neither lets you actually see — or delete — what's been learned about you.

Priver Mneme takes a different path. Memory is stored as a structured event graph inside the frame, recalled by a purpose-built co-processor in milliseconds, and shielded by a privacy boundary built into the silicon itself.

How it remembers

Memory that grows, fades, and strengthens — like yours.

Priver Mneme doesn't record video. It distills what matters into typed event nodes — people, places, actions, moments — connected by associative links. Each link carries a confidence value and a weight that grows when patterns repeat and quietly decays when they don't.

01 / NODES & LINKS

A graph, not a recording

Entities, actions, states, moments and the reasons between them are stored as a typed event graph — each link weighted by confidence, recurrence and source reliability. Compact, structured, and built to be read by a human.

02 / GROW · DECAY · REINFORCE

It learns without retraining

Repeated patterns strengthen. Contradicted or unused ones decay toward a floor. Good outcomes reinforce the path that led to them — all on-device, with no neural network ever retrained on your data.

03 / RECALL & PREDICT

Surfaces before you ask

A dedicated traversal co-processor follows the strongest links from what you're looking at to recall the relevant memory in under 100ms — and pre-loads what it predicts comes next, before the moment fully arrives.

The privacy boundary

Raw camera and audio physically can't leave.

Privacy here isn't a setting or a promise in a policy. It's a wall etched into the silicon. A sensor-side chip turns what the camera sees into a normalized event structure, wipes the frame buffer, and the raw image crosses a one-way bus that has no return path.

  • BUS

    Unidirectional by construction

    The raw sensor bus only flows one way. There is no electrical path back to the application processor or the radio.

  • WIPE

    Frame buffer cleared on extraction

    Once an event is extracted, the dedicated ASIC clears the buffer holding the raw image. The original frame ceases to exist.

  • SHARE

    Only weights travel, never frames

    Between your devices, Priver Mneme syncs changes to link weights and confidence — never raw sensor data, never original input.

camera mic array eye + IMU raw frames → event extraction ASIC ⟲ buffer wiped HARDWARE PRIVACY BOUNDARY · ONE-WAY events ✓ raw blocked event-graph memory
Inside the frame

Four silicon blocks. One pair of glasses.

Everything that makes Priver Mneme work lives in the temples and bridge — separated by design, so that inference, safety, and your raw senses never share the same memory.

fwd camera + depth IR eye-tracking IMU + mic array capacitive temple co-processors + bone conduction
BLOCK A

Event-extraction ASIC

Sits beside the camera. Turns raw senses into events and enforces the privacy boundary.

BLOCK B

Graph-traversal co-processor

Sparse-graph fetch, link-weight math and an on-die node cache. Recall, not generation.

BLOCK C

Reality-check watchdog

An independent SoC on its own clock and power rail. Verifies before anything is shown.

BLOCK D

Foveated waveguide display

Renders crisp detail where you look, dimmer at the edges — anchored to your gaze.

Safety, separated

Nothing is shown until reality agrees.

A recall is only a candidate. Before any overlay, alert or haptic reaches you, an independent watchdog — with its own clock domain and power rail — checks it against what the sensors see right now. If memory and reality disagree, it doesn't render.

recall
candidate
reality-check
watchdog
validate vs.
live sensors
authorize
render ✓
Your memory, your rules

The first wearable memory you can read, edit, and erase.

Because memory is a structured graph instead of buried model weights, you can open it, change one fact, or delete it for good — and have that deletion follow you across every device.

Edit by hand or by glance

Touch the temple, dwell with your eyes, or gesture in view to correct a node or link — instantly, without retraining a thing.

Inspect on the desktop

Export a portable knowledge package and open your memory as a graph, timeline, heatmap or map. Branch edits, preview the effect, then commit.

Delete that actually deletes

Erasure writes a tombstone so the fact can't return, and sends correction signals that unwind its influence everywhere it spread.

Federated, trust-scaled

Your devices share a single memory through a portable vault. Updates from peers are weighted by trust; access can be granted and revoked, one system at a time.

Specification · Priver Mneme

The short technical sheet.

Recall latency
< 100 ms, on-device — no cloud round-trip, no on-device generative model
Memory store
Persistent typed event-graph in non-volatile memory — within the frame
Privacy
Hardware-enforced boundary — unidirectional sensor bus + frame-buffer clear ASIC
Recall engine
Graph-traversal accelerator co-processor — sparse-graph fetch, on-die node cache, forward-traversal preload
Safety
Independent reality-check watchdog SoC — separate clock domain & power rail
Display
Optical waveguide, foveated — gaze-anchored overlays, high-res foveal region
Sensing
Fwd camera · depth · IR eye-tracking · IMU · beamforming mic array
Input
Capacitive temple touch · dwell-confirm gaze · forward-camera gesture
Output
Waveguide overlay · bone-conduction audio · haptic
Connectivity
Ultra-wideband · Bluetooth Low Energy — companion device & desktop data interface
Learning
Grow · decay · reinforce — no neural-network retraining on personal data
Portability
Exportable knowledge package — inspect, edit, re-import with reconciliation
Now reserving · first allocation

Put your memory where it belongs.

On your face, in your hands, and nowhere else. Reserve a pair and be among the first to wear a memory that stays yours.