flashbulb-amd: zero-token agent memory on AMD

Every shipped agent-memory system, like Mem0, Zep, and Letta, pays two taxes on the write path. The first is a token meter. Each turn, they call an LLM to pull out facts and decide whether to add, update, or ignore them, and that per-turn cost climbs with usage (roughly $5,000 to $30,000 a month at 10M turns). The second tax is data egress you can't switch off, because that same call sends your agent's memory to a cloud API. For a bank, a hospital, or a government agency, that isn't a line item. It's a compliance blocker. The teams who most need durable agent memory often can't let it leave the building. flashbulb-amd drops both. A small Titans-style neural memory (about 2.8M parameters), trained as it runs, gives each new observation a learned "surprise" score: how badly the current memory predicted it. High surprise, it gets stored, like a flashbulb memory. Low surprise, it gets skipped. That score is a 5 ms tensor op, not an LLM call, so it costs zero tokens, and the network runs locally, so nothing is sent anywhere. Forgetting comes from weight decay: when a new fact contradicts an old one, the stale fact fades. The design is deliberately hybrid. The neural memory only decides what is worth keeping; the actual text is recalled from a plain SQLite and embeddings store. The whole stack runs on one AMD Instinct MI300X. The surprise gate, the memory store, and a Gemma 4 model (26B-A4B) all sit on a single 192 GB accelerator, ROCm throughout, which is what lets you self-host with no data leaving the box. In our benchmark that came out to 0 write-path tokens versus 2,055 for the LLM-extraction baseline over 30 turns, about 37 times faster per decision, and 96.7% gate accuracy. The business model is open core: free and MIT-licensed for developers who self-host, paid for enterprises that need air-gapped deployment, compliance, and support. A live dashboard shows the surprise signal, the running token counter, and facts being forgotten as you type. Have fun scoring!