VoidOrigin is a dev lab. We threat-model security, build AI systems, and ship production apps. Security work stays private with the client; the AI research is open on GitHub.
Fig. 01From kongclaw/src/acan.ts. Real code.
We are building a cognitive memory model for AI agents. Graph-backed, locally embedded, self-correcting. It decays stale knowledge, links corrections to the beliefs they replace, and finds what you forgot to ask about. Not another RAG. 98.2% Recall@5 on LongMemEval, zero cloud dependency.
LongMemEval
KongCode suite
emit to NT8
MIT licensed
Selected work
Benchmark
Recall @ 5 · LongMemEval · cross-attention reranked · zero API calls
The memory layer behind KongClaw beats most published open-source baselines on LongMemEval, and it does it without a hosted vector DB or retrieval API. Everything runs locally: graph vector search generates candidates, then an on-box attention model reranks them and auto-trains from your retrieval outcomes.
What we do
Threat modeling. Audits. Production hardening.
Memory, agents, and retrieval systems.
Full-stack builds and real-time infra.
Advisory work is billed by the hour, builds run on monthly retainer, and audits are scoped per project. We can share a portfolio and named references for qualified inquiries, and we'll sign an NDA on request.
About the founder
Real engineering.
Real output.
Khris Smith founded VoidOrigin in 2023 after a decade in production engineering and three years on AI and cognitive memory research. He still runs the lab himself, and everything shipped is built in-house.
Set in
Inter & JetBrains Mono
Built with
Next.js · Tailwind · FastAPI
Served via
Cloudflare & AWS