Hooks and adapters that give agents memory, retrieval that stays silent until it's relevant,
and desktop tooling that makes the whole loop visible. Spec first, small verifiable changes, tight feedback loops.
Okryptos 归藏 — a latent knowledge base for AI-agent hooks · website
Knowledge rests silent until an agent hook calls it into context. Ten AI assistants — Kimi Code, Claude Code/CodePilot, Codex, opencode, Pi, ZCode, Reasonix, Qoder (CLI & IDE), DeepSeek Harness — share one per-project knowledge base through an extensible adapter architecture.
- Retrieval with discipline — hybrid keyword + vector search, per-channel admission (better none than noise), coverage gating for short colloquial queries, and an LLM relevance post-filter that only ever removes candidates, fail-open all the way down
- Local-first inference — embedding and intent filtering run on managed llama.cpp sidecars (Ollama or builtin profiles), zero API cost, knowledge never leaves the machine
- Workflow enforcement — a turn that changed code without writing a changelog gets blocked until it's fixed
- Ships as a single-binary Go CLI + web GUI + resident daemon; Windows installer, Linux deb / tar.gz
OkMeter — a desktop token-usage dock (C++20, D3D11/Direct2D)
A screen-edge meter for agent token spend: six forms × four materials, real background-refraction liquid glass via Windows Graphics Capture, cascading glass menus, and a back-panel settings UI — all on a self-rendered swapchain with spring animations and on-demand frame scheduling.
Multi-device sync — okserver (Gitea-backed sync service) + okdeploy, a one-click NAS deployment console; released as multi-arch images to GHCR and Docker Hub.
Go · C++ (Win32/D3D11/D2D/WGC) · Python · TypeScript · Shell · Inno Setup · SQLite · llama.cpp
- Agent memory: capture flows, retrieval admission, injection budgets, cross-turn cooling
- Local LLM profiles for retrieval-time filtering — smaller models, tighter gates
- Native desktop rendering: WGC capture pipelines, D2D material effects, animation schedulers
- Main development happens on a self-hosted Gitea — GitHub is the public mirror; releases publish to both
- This profile is intentionally lightweight — the interesting parts live in the repositories
