Docs Home
Mnemon Harness Design Note

生命周期控制平面

Mnemon 保存 State，声明 Intent，观察 Reality，并通过 Reconcile 对齐差距。实体不是模型本身；实体只是模型中的 profile。Execution Surfaces 负责把 Mnemon 的意图投影给宿主，并把宿主现实观察回来。

state intent reality reconcile

Lifecycle Control Plane

Mnemon keeps State, declares Intent, observes Reality, and Reconciles the gap. Entities are not the model itself; they are profiles inside the model. Execution Surfaces project Mnemon intent into the host and observe host reality back into Mnemon.

state intent reality reconcile

最小定义

不强行对齐外部系统概念，只保留 Mnemon 自己的一致结构。

Mnemon 保存 State，声明 Intent，观察 Reality，并通过 Reconcile 把 Reality 拉回 Intent，结果重新写入 State。

Core Model

这是 Mnemon 如何理解世界；Surface 不在这一层，而在执行层。

State durable truth .mnemon 中的 memory、skills、reports、proposals、audit、status。
Intent desired lifecycle Mnemon 声明希望系统呈现的生命周期形态。
Reality observed world 宿主、项目、工具、eval 和运行时当前真实发生的状态。
Reconcile alignment 比较 Intent 与 Reality，并把结果写回 State。

控制闭环

这个闭环是稳定内核；具体实体只是进入闭环的不同 profile。

State 持久上下文、控制状态、证据和治理记录。
Intent 期望的 lifecycle shape、policy、binding 或 proposal。
Reality 通过宿主 surface 观察到的文件、运行时、eval、工具状态。
Reconcile repair、validate、propose、review 或 no-op。
State status、report、proposal、audit 成为下一轮输入。

Entity Profiles

实体不是模型本身。实体只需要声明自己在 Core Model 中扮演什么 profile。

Template reusable definition 定义可复用能力，不一定被持续 reconcile。例：Loop。
Controlled spec / status 需要持续对齐 Intent 与 Reality。例：Binding、EvalRun、未来 Goal。
Surface reach reality 表达或触达宿主能力。例：HostCapability、Projection。
Evidence observed fact 来自 Reality 的证据，不是声明对象。例：Observation、runtime status。
Governance risk boundary 处理人审和审计。例：Proposal、Review、Audit。

当前实体映射

只列当前最需要理解的实体，保持页面轻量。

Entity Profile Role
Loop Template 可复用生命周期能力包，例如 memory、skill、eval。
Binding Controlled 把某个 Loop 绑定到某个 host；适合作为第一个完整 controlled object 样本。
HostCapability Surface 描述宿主可暴露的静态或动态能力。
Projection Surface 让 HostAgent 看见 Mnemon 的 Intent。
Observation Evidence 让 Mnemon 看见 HostAgent 的 Reality。
Proposal / Review / Audit Governance 当 Reconcile 无法安全自动完成时，保存提议、决策和不可变记录。

Execution Surfaces

这一层解释 Mnemon 如何触达宿主，不再和 Core Model 混在一起。

Projection

静态方向：把 Intent 渲染成 host-readable view。

.codex/skills .claude/hooks config generated docs manifest

Observation

动态方向：把 Reality 转化为 status、evidence 或 proposal 的输入。

Codex appserver session API eval endpoint tool status runtime errors

Memory-loop 给出的证据

memory 不是全部答案，但它验证了 Mnemon 的方法论。

模式 常见做法 Mnemon 做法 抽象结果
Memory 外部 memory service、vector DB、retrieval API。 通过 hook + skill + .mnemon state 进入 prime、remind、nudge、compact。 memory 从存储能力变成 lifecycle-native capability。
Self-improvement 独立实验平台或单 repo overnight loop。 通过 hook + skill + daemon + HostCapability 进入可治理 project evolution。 autoresearch-like loop 变成可治理生命周期能力。

与 autoresearch 的关系

参考的是自改进循环的纪律，不是复制一个实验平台。

Autoresearch

  • 极简自主实验循环。
  • 一个主要可变文件、一个固定时间预算、一个指标。
  • 核心决策是 keep 或 discard。

Mnemon

  • 面向持久自主改进循环的生命周期控制平面。
  • 用 Entity Profiles 让不同实体轻量进入同一个控制模型。
  • 把决策扩展为 repair、validate、propose、review、audit。

演进层级

保持轻量：先定义 profile，再扩展 surface，最后进入自演进治理。

Level 1: Profiles 每个实体声明 profile，不急于变成完整资源对象。
Level 2: Projection 把 Intent 投影给 HostAgent。
Level 3: Observation 通过 appserver、eval、tool status 观察 Reality。
Level 4: Governance 让 AI 产生 patch、report、proposal，由 review gate 控制风险。

Minimal Definition

Do not force-fit external concepts. Keep Mnemon's own structure consistent.

Mnemon keeps State, declares Intent, observes Reality, and uses Reconcile to pull Reality back toward Intent, writing the result into State.

Core Model

This is how Mnemon understands the world. Surface belongs to the execution layer, not this layer.

State durable truth Memory, skills, reports, proposals, audit, and status under .mnemon.
Intent desired lifecycle The lifecycle shape Mnemon wants the system to present.
Reality observed world The current real state of the host, project, tools, evals, and runtime.
Reconcile alignment Compare Intent with Reality, then write outcomes back into State.

Control Loop

This loop is the stable kernel. Concrete entities enter it through different profiles.

State Durable context, control state, evidence, and governance records.
Intent Expected lifecycle shape, policy, binding, or proposal.
Reality Files, runtime, evals, and tool status observed through host surfaces.
Reconcile repair, validate, propose, review, or no-op.
State Status, reports, proposals, and audit become next input.

Entity Profiles

Entities are not the model itself. Each entity only declares its profile inside the Core Model.

Template reusable definition Defines reusable capability and is not necessarily reconciled. Example: Loop.
Controlled spec / status Needs ongoing alignment of Intent and Reality. Examples: Binding, EvalRun, future Goal.
Surface reach reality Expresses or reaches host capability. Examples: HostCapability, Projection.
Evidence observed fact Evidence from Reality, not a declarative object. Examples: Observation, runtime status.
Governance risk boundary Handles review and audit. Examples: Proposal, Review, Audit.

Current Entities

Only list entities that must be understood now, keeping the page lightweight.

Entity Profile Role
Loop Template Reusable lifecycle capability package such as memory, skill, eval.
Binding Controlled Binds one Loop to one host; suitable as the first full controlled object sample.
HostCapability Surface Describes static or dynamic capabilities a host can expose.
Projection Surface Lets the HostAgent see Mnemon's Intent.
Observation Evidence Lets Mnemon see the HostAgent's Reality.
Proposal / Review / Audit Governance Stores proposals, decisions, and immutable records when Reconcile cannot safely complete automatically.

Execution Surfaces

This layer explains how Mnemon reaches the host without mixing it into the Core Model.

Projection

Static direction: render Intent into a host-readable view.

.codex/skills .claude/hooks config generated docs manifest

Observation

Dynamic direction: turn Reality into status, evidence, or proposal input.

Codex appserver session API eval endpoint tool status runtime errors

What Memory-loop Proved

memory is not the whole answer, but it validated Mnemon's method.

Pattern Common Path Mnemon Path Abstraction
Memory External memory service, vector DB, retrieval API. hook + skill + .mnemon state across prime, remind, nudge, and compact. Memory becomes a lifecycle-native capability.
Self-improvement Standalone experiment platform or one-repo overnight loop. hook + skill + daemon + HostCapability for governable project evolution. Autoresearch-like loops become governable lifecycle capabilities.

Relation To Autoresearch

Borrow the discipline of self-improving loops, not the shape of an experiment platform.

Autoresearch

  • A minimal autonomous experimentation loop.
  • One main mutable file, one fixed time budget, one metric.
  • The core decision is keep or discard.

Mnemon

  • A lifecycle control plane for durable autonomous improvement loops.
  • Uses Entity Profiles to keep different entities inside one lightweight model.
  • Extends decisions into repair, validate, propose, review, and audit.

Evolution Levels

Stay lightweight: define profiles, expand surfaces, then move toward self-evolution governance.

Level 1: Profiles Every entity declares a profile before becoming a full resource object.
Level 2: Projection Project Intent into the HostAgent.
Level 3: Observation Observe Reality through appserver, eval, and tool status.
Level 4: Governance Let AI produce patches, reports, and proposals while review gates control risk.