FRIDAY · MAY 29, 2026Open for work
VoidOrigin · est. 2023

>we break it.>we write it.>we ship it.

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.

kongclaw/src/acan.tsmaster
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
TypeScript·utf-8·LFtyping

Fig. 01From kongclaw/src/acan.ts. Real code.

New/Cognitive memory model

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.

RetrievalGraph vector search + ACAN cross-attention rerank
DecayKanerva-Tulving stability decay, floor 0.15
CorrectionSupersede edges, 0.4x decay, both queryable
CostZero external API calls on the read path
Recall @ 5
98.2%

LongMemEval

Tests green
1014

KongCode suite

Signal push
<250ms

emit to NT8

OSS shipped
6+

MIT licensed

II.

Benchmark

Memory recall. Cross-attention reranked.

98.2%

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.

  • RetrievalGraph vector search, k=128 candidates
  • RerankACAN cross-attention, on-box, ~130K params
  • TrainingPure TS SGD on retrieval outcomes, ~30s
  • CostZero external API calls on the read path
III.

What we do

We work in three areas. Security, AI, Apps.

Security

Threat modeling. Audits. Production hardening.

  • Threat model and attack-path walkthroughs
  • Auth, session, and token lifecycle audits
  • Web + API pen testing
  • Infrastructure review and remediation

AI

Memory, agents, and retrieval systems.

  • Agent architecture + memory review
  • Cognitive retrieval systems (graph + rerank)
  • LLM integration with cost + latency surgery
  • Claude Code / MCP plugin development

Applications

Full-stack builds and real-time infra.

  • Next.js + FastAPI production stacks
  • WebSocket delivery + real-time plumbing
  • Auth, billing, and admin ops out of the box
  • Observability, audit logging, retention
Dev retainer$2,250 / mo

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