I'm an Agent systems builder working on local-first runtimes, MCP/A2A tooling, and auditable memory.
我在构建真正能长期生活在用户身边的 AI Agent：本地优先、可审计、可组合，并且尊重人的控制权。
I care less about making a model look impressive and more about making Agent systems reliable: structured memory, traceable state, composable tools, and automation that keeps people in control.
| System | What I'm exploring |
|---|---|
| 🫀 Vital Agent Sync | A local-first Apple Health connector that exposes user-controlled, scoped context to MCP-compatible Agents. |
| 🧠 MemTable | A structured ledger for Agents: confirmed records, timestamps, schemas, aggregation, and source tracing — not another vector-memory wrapper. |
| 📚 Oh My Notion | A local-first Notion Agent with hybrid retrieval, skill routing, grounded answers, and inspectable memory. |
Memory recalls what happened. A ledger computes what changed.
| Project | Merged contribution |
|---|---|
| First Tree | Fixed session-operation queue cleanup while preserving per-session FIFO ordering across WebSocket close, with race-regression coverage. |
| AGNTCY Directory | Added a CountRecords gRPC API with distinct-record counting, generated bindings, and database/controller tests. |
| Atomic Agents | Added project-instruction support for multiple AI coding assistants beyond a single vendor workflow. |
- A2A JavaScript SDK #178 — assigned to build public-API compatibility checks for breaking-change detection.
- Agentgateway #673 — scope confirmed for protocol-aware ALPN defaults.
- OpenHands #15583 — approved to implement slash-command parity, starting with
/help.
local over opaque
inspectable over magical
composable over monolithic
consent-aware over always-on
useful over impressive
Blog · X / Twitter · GitHub
Building Agents that grow useful without growing beyond their users' control.