本篇正文写于命名裁决之前,仍在使用旧名:
本境 在此指长期存储(现为 学籍 Xueji,本境已还给环境层)、
ActionParity 与 OriginBus 在此当部件名用(二者是商标,
对应的技术名是 影核 Action Kernel 与 北桥 Northbridge),
观察器本象现已拆分为 取象 Quxiang。
裁决与理由见 命名裁决 · EN。
This document predates the naming decision and still uses the retired names. The decision is published rather than quietly applied, because renaming the prose is a separate editorial change that has not been made yet.
RFC-0000 — 2Origin Computer Architecture
Status: Draft · Request for Comments Version: 0.1 Date: 2026-08-08 Author: U-King
We don't build models. We want to build an AI computer. The model is the CPU — but a CPU is never a computer.
Principle: Model is replaceable. State survives. Actions are portable. Learning compounds.
1. Goal
This architecture defines a persistent computer architecture for AI agents.
It does not define models. It does not replace MCP, A2A, or any specific Harness (agent runtime). It defines the layer above all of them:
How does an AI computer obtain world state, persist long-term state, execute real-world actions, and learn from every job — independent of model and independent of Harness?
A conforming machine must allow DeepSeek, GPT, Claude, Kimi, or any future model to run as a swappable compute unit — without losing the environment, projects, skills, and experience the machine has accumulated.
2. Why an "AI computer" and not "another agent"
Today's agent stack already has good parts:
- Model (DeepSeek / GPT / Claude) — CPU
- Context Window — RAM
- MCP / A2A — device bus / inter-process communication
- Harness (Claude Code / Codex / OpenHands) — kernel / scheduler
But they don't assemble into one computer, because the motherboard, disk, drivers, and operating system are missing.
The universal pain point:
Taught today, forgotten tomorrow. Learned on this task, restarting from grade one on the next Session.
The reason: agents persist chat transcripts (shadows) instead of object state (origin). State is "what the world is right now"; a transcript is "who said what." An AI always sees the shadow, never the object itself.
3. Terminology & Layers
3.1 Architecture
| Layer | PC analogue | In this system | Responsibility |
|---|---|---|---|
| Intelligence | CPU | Model (swappable) | reasoning, planning |
| Working memory | RAM | Context Window | current thinking |
| Long-term storage | SSD + FS | Benjing 本境 | everything this AI learned |
| World representation | GPU / scene graph | Benxiang 本象 | world → AI-computable objects |
| Action I/O | Southbridge + drivers | ActionParity 影核 | changing the world (GUI/CLI/API/device) |
| High-speed bus | Northbridge | OriginBus | compiling relevant state into Context |
| Kernel / boot | BIOS + Kernel | Harness (swappable) | scheduling, context, tools |
| Self-learning | system service | Academy 学堂 | distilling experience into credentials |
| The machine | PC | U-King | first reference implementation |
3.2 One-liners
Benxiang saves the world. Benjing saves growth. ActionParity changes the world. The Northbridge knows (Know What). The Southbridge acts (Do What).
4. The full task loop
A completed task must traverse the full loop:
Observe → Think → Act → Verify → Learn
① BOOT Harness starts → mount Benjing → load project/environment
② OBSERVE Benxiang reads world state → Northbridge filters → into Context
③ THINK Model reasons → produces a plan
④ ACT Harness schedules → Southbridge → ActionParity → the world
⑤ VERIFY Observe the world again → confirm it truly succeeded (not exit code = 0)
⑥ LEARN Academy analyzes the trace → experience/SOP/skill → back to Benjing
⑦ NEXT BOOT The AI understands this machine better than last time
Constraints
- Verify is not "tool returned success." Success must be confirmed by re-observing real state (State Diff).
- Learning must be candidate → verified. One success is not a permanent truth.
candidate → reviewed → verified → deprecated / superseded
5. State format (minimal Benxiang)
All significant objects use one Envelope:
{
"spec": "2origin/0.1",
"kind": "task.origin",
"id": "2o:task:001",
"version": 1,
"scope": "project:demo",
"created_at": "ISO8601",
"updated_at": "ISO8601",
"goal": "...",
"current_state": "...",
"facts": [{ "claim": "...", "verified": true, "source": "..." }],
"decisions": [{ "what": "...", "why": "..." }],
"actions": [{ "verb": "...", "status": "done" }],
"artifacts": ["..."],
"verification": "...",
"next_steps": ["..."],
"learnings": [{ "lesson": "...", "confidence": 0.9, "status": "candidate" }]
}
Iron rules:
- facts must be verified. Unverified is a hypothesis, not a fact.
- Persist State + Facts, not chat transcripts. Transcripts are shadows; Benxiang is the object itself.
- Learning starts as candidate, becomes verified. One success is not a permanent truth.
6. Learning loop (Academy)
complete task → record trajectory → verify result → distill success/failure
→ abstract SOP → decide if long-term → write to Benjing → generate/update Skill → auto-invoke next time
The Academy produces a portable AI Transcript. Models may be swapped; the transcript persists.
7. Conformance
An implementation is 2Origin Compatible only when it satisfies all of:
- Cross-Session state: closing and reopening a session resumes via State, not transcript replay.
- Cross-Harness: switching harness (Claude Code ↔ Codex ↔ OpenHands) keeps the state usable — no re-asking.
- Cross-Model: switching models keeps the state usable — credentials do not reset.
- Portable actions: the same action can be executed by different drivers (API/CLI/GUI/device).
- Verifiable results: outcomes confirmed by observing real state, not just exit codes.
- No auto-permanent learning: experience must pass candidate → verified.
- Auditable: all critical actions, promotions, and policy changes are recorded.
8. Verified evidence (2026-08-08, first run)
First-run pass = signal the architecture is right (not a benchmark-tuned result).
A — Cross-Session ✅
A fresh, independent Claude Code session (claude -p, zero prior conversation) received the task.origin.json summary automatically via a SessionStart hook and correctly reported the task title and goal — no user re-explanation needed.
B — Cross-Harness ✅
Claude Code did STEP-A (distilled 12 facts into extracted-facts.md, updated state). Codex (codex exec) continued STEP-B from only task.origin.json + facts with zero re-asking. Session logs confirm Codex never asked "what is the task?"
C — Southbridge write access: SOLVED ✅ (was a known limitation)
What cross-harness really lacked was not the state format (Benxiang solves that) but a unified write authorization in the OriginBus Trust layer — the Southbridge's job. A Southbridge MCP server (southbridge_write) provides exactly that: a whitelisted, audited write path for headless harnesses.
End-to-end verified (2026-08-08): Codex (codex exec --approve-for-me) called southbridge_write to persist demo/southbridge-e2e.md → OK: wrote 25 bytes. The audit log records the action (actor, action, target, bytes, timestamp). Whitelist check denied a path-traversal attempt (evil/../shadow.txt) earlier — the Trust layer blocks what it should.
Why other approaches failed: Codex's read-only sandbox blocks shell writes; -s workspace-write doesn't expose apply_patch; plain codex exec cancels MCP calls for lack of an approval channel. The Southbridge's MCP tool + --approve-for-me closes the loop. The earlier blocker (expired Codex login token) was an environment issue, now resolved.
9. Relation to existing standards
- Does not replace MCP (tool/data connection), A2A (agent interop), or Harness (runtime).
- This architecture is one level higher: how to assemble them into a computer that works long-term and gets better with use.
- Implementations may use MCP / A2A / any harness as transport — no conflict.
10. Open questions (Request for Comments)
- Benjing layering: SQLite fact source + file snapshots + optional vector index — sufficient?
- What standard interface should the Northbridge Context compiler (relevance + token budget) expose?
- Minimal safe model for Southbridge write authorization (Trust/Approval)?
- Which Academy promotions should be automatic vs. human-reviewed?
Related RFC: RFC-0005 — Benjing v0.2 (content-hash optimistic lock / verifiable sources / actor provenance / bundle compilation) · RFC-0006 — Northbridge & Southbridge (context.request→bundle / risk tiers + approval)
License: Apache-2.0 Trademark: U-King and 2Origin are trademarks. See TRADEMARKS.