Perseus Vault — Encrypted Agent Memory for AMD

Perseus Vault is encrypted, persistent memory for AI agents. It ships as a single ~8 MB Rust binary exposing 56 MCP tools plus five PyPI adapters — LangChain, LlamaIndex, Haystack, CrewAI, Pydantic AI. This is a shipping product, not a weekend build: v2.20, 35 releases. The money shot: MI300X serving Qwen2.5-72B bf16 on vLLM/ROCm holds 15.3 concurrent agents at $0.143/agent-hour, 658 output tok/s. With the GPU saturated, the memory layer steals only ±0.6% recall (6-run median) — near-zero inference cost. Measured cross-vendor: 11.7× cheaper per agent than 2×H100, 3.0× cheaper than 2×A100 80GB, and benchmarked against 8×A100 40GB too — every row measured, none projected. Keeping memory off the accelerator is AMD's killer advantage for stateful agent fleets. Local-first, no API key required: Gemma 3 4B on a Ryzen 7 9800X3D recalls in 0.21ms at ~13 tok/s for $0. Air-gapped deployments lose nothing. Live demo: https://amd-demo.perseus.observer — teach a fact, recall it in a new session, watch an open-weight LLM answer from memory alone. Code: https://github.com/tcconnally/perseus-amd-act-ii