Perseus Vault × AMD Instinct MI300X

Encrypted, local-first agent memory — kept off the GPU. Recall runs on the host CPU; the agent's answer is generated live by an open-weight LLM via the Fireworks AI API.
Judges: tap an example below to teach a fact, then hit Recall — and read the live LLM answer grounded only in what was recalled.

⚠︎ Honesty: recall & footprint are measured live on this CPU (0 bytes of GPU HBM). The agent's answer is generated live by an open-weight LLM via Fireworks AI — the hackathon's inference partner, now partnering with AMD to serve on Instinct; no serving API attests which accelerator handles a request, so we don't claim one. (Daily budget cap; falls back to a labelled memory-grounded composition.) The cost table below is measured on real rented hardware — one MI300X and one 2×H100 pair (Qwen2.5-72B, vLLM); only the A100 row is a projection.

Session 1 · teach the agent

Store durable facts. Tap an example or type your own.

active memories: 0

nothing stored yet.

Session 2 · new session, recall

clears the on-screen context — the vault persists. Then recall.

start a new session, then ask something you taught.

Why AMD: one card, many agents measured · MI300X + 2×H100

serving one 72B model; memory stays on CPU (0 HBM).

Accelerator	HBM	Cards	Agents	$/GPU-hr	$/agent-hr	source

11.7× cheaper per agent-hour than 2×H100 — measured, not projected. A single H100 can't load the model at all; the MI300X holds it on one card with KV headroom to spare.

Same recall design as the shipping product (CPU reference implementation): perseus-vault · submission repo perseus-amd-act-ii · video · MIT. Per-visitor sandbox, auto-resets when idle.