✅ Cross-session
Close the window, resume the task. A fresh, independent Claude Code session (claude -p, zero prior conversation) received the task.origin.json summary via a SessionStart hook and reported the task title and goal correctly.
2Origin Computer Architecture · Draft · RFC v0.1 · Apache-2.0
The model is the CPU — but a CPU is never a computer.
模型是 CPU，但 CPU 从来不等于一台计算机。
Model is replaceable · State survives · Actions are portable · Learning compounds
Taught today, forgotten tomorrow.
Learned on this task, restarting from grade one on the next session.
Today's agent stack has good parts: models (CPU), context windows (RAM), MCP (buses), harnesses (kernels). But they don't assemble into one computer — the motherboard, the disk, the drivers and the OS are missing.
The root cause is plain: agents persist chat transcripts (shadows), not object state (origin).
| Layer | PC analogue | This system | Responsibility |
|---|---|---|---|
| Intelligence | CPU | Model (swappable) | reasoning |
| Working memory | RAM | Context Window | current thinking |
| Long-term storage | SSD | Xueji 学籍 | everything the AI learned |
| Sensing | sensor hub | Quxiang 取象 | reads the world — and never receives an expectation |
| World representation | GPU | Benxiang 本象 | world → AI-computable objects |
| Machine environment | CMOS / firmware | Benjing 本境 | what this machine has and can run |
| Action I/O | Southbridge I/O, out | Action Kernel 影核 | changing the world |
| High-speed bus | Northbridge | Northbridge 北桥 | state → Context |
| Kernel | BIOS + Kernel | Harness (swappable) | scheduling |
| Self-learning | system service | Academy 学堂 | experience → credentials |
| The machine | PC | U-King | first reference implementation |
Quxiang sees the world. Benxiang saves it. Xueji keeps the record. The Action Kernel changes it.
The Northbridge knows. The Southbridge acts.
Observe → Think → Act → Verify → Learn
A first-run pass is a signal that the architecture is right — not a benchmark-tuned result. The third card below did not pass, and it is on this page anyway.
Close the window, resume the task. A fresh, independent Claude Code session (claude -p, zero prior conversation) received the task.origin.json summary via a SessionStart hook and reported the task title and goal correctly.
Swap the engine, keep the credentials. Claude Code did half the task (distilled 12 facts); Codex (codex exec) continued from task.origin.json + facts alone with zero re-asking. Session logs confirm it.
Codex's sandbox is read-only. The fix is a Southbridge write action — an audited MCP tool (whitelist + audit log) injected into the harness. Prototype self-tests pass; the final end-to-end run was blocked by an expired login token — an environment issue, not an architectural one.
“World-first long-term AI computer prototype” is the initiator's falsifiable public claim, not a third-party certification. Earlier comparable public implementations are welcome as evidence; verified counterexamples will change the wording in public.
Small company, large appetite. Code, experiments, failures and retractions are all public — we don't ask you to believe first.