# Sandisk research hardware update

Date: 2026-06-23

Decision: the SanDisk research changes the hardware **positioning and roadmap**, but it does not invalidate the first Witness Vault build.

The first physical build should remain the Raspberry Pi Zero 2 W Witness Vault because it is the fastest credible way to make Enigma memory custody physical, inspectable, and demoable. The research adds a second hardware track: an NVMe-backed **Enigma Memory Node** for actual edge-cache and retrieval-locality work after the proof loop is stable.

## What the research changed

### 1. The box is not an AI accelerator

SanDisk became important because AI systems need memory hierarchy, bandwidth, persistence, locality, and power efficiency. That does not mean Enigma should build a tiny computer that claims to outperform a PC.

The Witness Vault must be framed as:

> A physical AI memory custody and witness appliance.

Not:

> A faster AI computer.

The PC wins raw compute. The Witness Vault wins trust, portability, isolation, visible status, local proof export, and a clean boundary.

### 2. Physical memory is the metaphor

SanDisk made memory tangible. Enigma needs the same tangible feeling for AI memory.

The physical product should make these abstract Enigma concepts visible:

- Passport ownership,
- local vault custody,
- receipt creation,
- deletion/tombstone evidence,
- witness checkpointing,
- offline verification,
- provider boundary honesty.

The OLED/NFC/printed enclosure matter because they turn a protocol into an object a user can point to.

### 3. The first build should prioritize proof over throughput

The current $50-$100 build uses:

- Raspberry Pi Zero 2 W,
- microSD local vault storage,
- ATECC608A secure-element sidecar,
- SSD1306 OLED,
- optional PN532 NFC,
- PETG/ASA printed enclosure.

That remains the right first build because Enigma already has a Node/CLI/MCP/verifier workflow. A Pi can run the software now. An ESP32 or NFC-only tag would force an embedded rewrite and would not host the local vault.

### 4. The second hardware product should be memory-first

The SanDisk research does change the next hardware target.

After the Witness Vault proves the custody loop, build:

> **Enigma Memory Node** — an NVMe-backed local AI memory cache and witness node.

Purpose:

- store larger encrypted memory capsules,
- host local vector indexes,
- cache hot memory close to agents,
- serve MCP/local-gateway context faster than cloud retrieval,
- checkpoint roots to the witness network,
- act as a home or small-office memory node.

Reference direction:

- Raspberry Pi 5 or Compute Module class host,
- NVMe HAT or M.2 carrier,
- 256GB–2TB NVMe SSD,
- secure-element or TPM sidecar,
- OLED/status LED,
- optional NFC/QR pairing,
- printed or CNC enclosure with accessible storage service panel.

Budget note:

A real NVMe node usually exceeds the original $50-$100 Witness Vault target. Do not squeeze it into the first budget by weakening the proof appliance. Keep V1 cheap and honest; make V2 faster and memory-centric.

### 5. Enterprise hardware should be a gateway, not a gadget

For enterprises, the physical device should become:

> **Enigma Enterprise Memory Gateway** — a rack/mini-PC/VPC appliance that keeps company AI memory under policy before it reaches models.

Purpose:

- BYOK/KMS,
- SSO/SCIM,
- SIEM export,
- policy receipts,
- data residency enforcement,
- local memory cache,
- model/provider allowlist,
- legal hold,
- witness checkpoints.

This is where performance and NVMe capacity matter most. The enterprise buyer does not need a cute object first. They need a policy-enforced memory boundary with proof export.

## What stays the same

The current Witness Vault remains correct for the first physical build.

Keep:

- Pi Zero 2 W as the V1 host,
- ATECC608A as the sidecar identity/key component,
- OLED as the operator trust surface,
- optional NFC as demo/pairing only,
- local vault bundle as canonical state,
- offline verifier as the proof authority,
- $50-$100 target for the first build,
- honest claim boundary.

Do not add claims that the V1 hardware cannot support.

## Physical design implications

### V1 Witness Vault design rules

- Keep the enclosure small and serviceable.
- Preserve antenna clearance for Wi-Fi/BLE and optional NFC.
- Keep status visible from the top face.
- Avoid conductive/carbon-filled filament near radios.
- Make the microSD accessible enough for maintenance.
- Keep fasteners accessible; do not glue the device shut.
- Do not market the printed case as tamper-proof.

### V1 demo flow

The demo should show custody, not benchmark performance:

1. Power on Witness Vault.
2. OLED shows `VAULT OK`.
3. AI client connects through MCP or local endpoint.
4. User retrieves a memory.
5. Enigma creates a context receipt.
6. User deletes/tombstones the memory.
7. Enigma verifies the export offline.
8. OLED shows `VERIFY OK` after the proof flow.
9. Optional NFC tap opens the receipt or device pairing page.

### V2 Memory Node design rules

- Add NVMe as the primary physical differentiator.
- Add thermal design before aesthetics.
- Add a serviceable storage bay.
- Add a clear distinction between plaintext local trusted mode and encrypted remote/witness mode.
- Keep chain coordination asynchronous; never put blockchain in the hot retrieval path.
- Benchmark retrieval latency, index load time, and encrypted capsule availability.

## Revised physical roadmap

| Stage | Product | Main job | Budget target | Build status |
| --- | --- | --- | ---: | --- |
| V1 | Witness Vault | Physical custody, local receipts, offline verification, demo trust surface | $50-$100 | Build this first |
| V1.5 | Witness Vault + removable capsule media | Portable encrypted proof/export bundle handoff | $65-$120 | Optional after V1 works |
| V2 | Enigma Memory Node / Memory Vault One | NVMe-backed local memory cache, vector index, home witness node | $210-$450 | Design after V1 proof loop |
| Enterprise | Enigma Memory Gateway | Policy-enforced company memory boundary with NVMe cache and audit exports | $550+ or BYOC/VPC | Sell after gateway pilot |

## Revised product thesis

Before research:

> Enigma needs a cool hardware vault.

After research:

> Enigma needs a memory hierarchy. The first object is the trust anchor. The second object is the fast local memory node. The enterprise object is the governed memory gateway.

The physical build now has a sharper role:

> **Witness Vault is the AI memory card people can see. Memory Node is the local AI memory cache people can feel. Gateway is the enterprise boundary people can audit.**

## Approved updated pitch

Use this:

> Enigma Witness Vault is not trying to beat your PC at compute. It gives your AI memory a physical home: local custody, hardware-backed device identity, visible receipt status, offline verification, and optional witness checkpoints. The next Enigma hardware tier adds NVMe-backed memory locality for larger private AI memory caches.

Do not use this:

> Enigma hardware makes AI 100x faster.

A future Memory Node can improve retrieval locality and privacy. It should only claim measured speedups after benchmarks exist.