SPL Agent Engine · 用户指南SPL Agent Engine · User Guide

用因果动力学重新定义 Agent 的状态演化机制：确定性人格、12 层连续记忆、可审计偏见。不再是给大模型打补丁，而是从底层范式解决概率偏移、短时记忆与 AI 偏见三大顽疾。A deterministic causal-dynamics engine that replaces probabilistic drift with reproducible state evolution: a fixed persona, 12-layer continuous memory, and auditable, tunable bias.

1. 产品简介1. Overview

SPL Agent Engine 是一套「决定论因果拓扑」驱动的拟人化智能体引擎。它把 AI 的内部状态建模为一套 12 层心理动力学系统：每一句话都会被解析为因果事件，事件沿因果拓扑网络确定性传播，实时改变 8 维情绪与信任、自尊、创伤等长期变量。SPL Agent Engine is a deterministic causal-topology engine for anthropomorphic agents. It models the agent’s internal state as a 12-layer psychodynamic system: every utterance is parsed into causal events that propagate deterministically through the network, updating 8-dimensional emotions and long-term variables such as trust, self-esteem and trauma.

2. 快速上手2. Quick Start

3. 核心机制3. Core Mechanics

3.1 8 维情绪流体3.1 8-Dimensional Emotion Fluid

快系统（秒级响应）实时表征情绪状态，与其他关键指标一起显示在演示面板：The fast system (second-level response) represents emotional state in real time, shown alongside key indicators on the demo panel:

3.2 12 层心理系统：连续动力学3.2 12-Layer Psyche: Continuous Dynamics

从快到慢、从表到深共 12 层，每层有独立时间常数（秒级 ~ 年级）。快系统积分驱动慢系统漂移，形成完整的时间尺度层级。Twelve layers from fast-shallow to slow-deep, each with its own time constant (seconds to years). Fast systems integrate and drift slow systems, forming a full temporal hierarchy.

3.3 因果链追踪：零概率 · 全追溯3.3 Causal Chain: Zero-Probability, Fully Traceable

「因果链」面板记录每一步状态变化及其完整传播路径。点击事件按钮后，8 维情绪、关键指标与因果链实时刷新，每一步变化都可查账。The Causal Chain panel records every state change and its full propagation path. Click an event button and the 8D emotions, key indicators and causal chain refresh in real time — every step is auditable.

f(stateₙ, event) → stateₙ₊₁   // 纯函数 · 无随机 · 可复现

f(stateₙ, event) → stateₙ₊₁   // pure function · deterministic · reproducible

子系统状态（身份 / 目标 / 价值观）由 feature/*.py 转译并实时显示；背叛、指责等事件会直接冲击 dignity · loyalty · honesty 等核心价值观；世界信念层「信任他人」会因负面经验下降、正面经验缓慢回升。Subsystem state (identity / goals / values) is compiled from feature/*.py and shown live; betrayal or blame directly hits values like dignity · loyalty · honesty; the “trust others” world belief falls on negative experience and recovers slowly on positive experience.

3.4 养成模式3.4 Raising Mode

同一只宠物，你怎么养，它就长成什么性格——每一步都能翻因果链查账。核心机制：Raise one pet and it grows the personality you cultivate — every step is an auditable causal chain. Core mechanics:

• 记忆珠 · 艾宾浩斯曲线Memory Beads · Ebbinghaus curve：「你们的约定」等记忆随时间自然变淡，符合艾宾浩斯遗忘曲线。Memories like “the promise we made” fade naturally over time along an Ebbinghaus forgetting curve.
• 照护强度Care intensity：调节 0–1，决定照护行为的力度与效果。Adjust 0–1 to control the strength and effect of care actions.
• 时间流转Time flow：快进 1 小时、睡一晚，观察慢变量的演化。Fast-forward 1 hour or sleep a night and watch slow variables evolve.
• 存档 / 读档 / 重置Save / Load / Reset：随时保存成长状态，或重置重新开始。Persist progress anytime, or reset to start over.

3.5 范式对比：造钟 vs 雇报时员3.5 Clock vs Clock-Watcher

同一个目标——让 AI 有「人心」——两条路做法完全相反：Two radically different paths to one goal — giving AI a “heart”:

为什么 LLM 原生 Agent 解决不了三个问题：不是调参问题，是底层范式问题——概率生成模型的本质缺陷无法通过 prompt engineering 或 RAG 根治。Why native LLM agents can’t solve the three problems: it’s not a tuning issue but a foundational paradigm issue — the inherent flaws of probabilistic generative models cannot be cured by prompt engineering or RAG.

• 黑箱 · 不可控：每轮对话都是概率采样，同输入不同输出；长期运行后人格、立场、知识一致性逐步漂移，无法锚定。Black box · uncontrollable: every turn is a probabilistic sample; Drift grows over long runs; persona and knowledge cannot be anchored.
• 转身就忘：上下文窗口有限，超出即遗忘；RAG 是「查资料」不是真正的状态连续——创伤、信任、心境等慢变量无法持久化。Forgets instantly: bounded context window; RAG is lookup, not state continuity — slow variables like trauma and trust can’t persist.
• 有偏见改不了：训练数据中的偏见被概率化放大，来源不可追溯，只能靠 RLHF 事后压制。Untunable bias: training-data bias is amplified probabilistically; untraceable, only suppressed post-hoc by RLHF.

3.6 偏见显式化：看得见，能调能关3.6 Explicit Bias: Visible, Tunnable, Zeroable

所有认知偏见（乐观偏见、确认偏见…）都是可配置的显式参数：来源可追溯，大小可精确调控，甚至可以设为零实现完全理性。责任闭环锚定到具体参数节点，而非黑箱概率。All cognitive biases (optimism, confirmation…) are explicit configurable parameters: traceable source, precisely tunable, and can be zeroed for full rationality. Accountability anchors to concrete parameter nodes, not black-box probabilities.

bias = optimism + confirmation   // 显式 · 可追溯 · 可归零

bias = optimism + confirmation   // explicit · traceable · zeroable

心理参数调节面板提供预设：乐观人格、偏执人格、创伤后、缺觉状态，可手动精细调整。The psychological parameter panel ships presets — optimistic, paranoid, post-trauma, sleep-deprived — and supports fine manual tuning.

3.7 API 集成：3 行代码接入3.7 API Integration in 3 Lines

引擎纯计算、无 IO，可嵌入任何运行时（Web / Node / 服务端）。工作台顶部「接入文档」章节展示完整 API 用法，支持厂商选择与大模型驱动的回复切换。The engine is pure computation with no IO and embeds in any runtime (Web / Node / server). The “Docs” section in the workbench nav shows full API usage, including provider selection and LLM-driven replies.

4. 应用场景4. Scenarios

选择一个业务场景，触发事件序列，观察因果链如何确定性地传播。所有状态变化 100% 可追溯。Pick a business scenario, fire an event sequence, and watch causal chains propagate deterministically. Every state change is 100% traceable.

5. 常见问题5. FAQ

Q1：这个引擎需要联网 / API Key 吗？Q1: Does it need network / an API key?

纯前端运行，因果引擎本体零 IO、零外部依赖；API Key 仅用于可选的大模型驱动回复。数据不出本机。Runs fully in-browser; the causal core has zero IO and zero external dependencies. An API key is only needed for the optional LLM-driven replies. No data leaves your device.

Q2：人格会漂移吗？Q2: Does the persona drift?

不会。所有状态变化都是确定性函数驱动，相同输入永远相同输出，长期运行零漂移、可复现、可测试。No. Every state change is a deterministic function — same input always yields the same output, with zero drift over time. Reproducible and testable.

Q3：为什么长对话后它还记得「之前的约定」？Q3: Why does it still remember earlier promises after long chats?

记忆不是存在上下文窗口里，而是存在 12 层连续动力学中：慢系统（信任、自尊、创伤）有月年级时间常数，记忆珠按艾宾浩斯曲线自然衰减而非截断。Memory lives not in a context window but in the 12-layer continuous dynamics: slow systems (trust, self-esteem, trauma) run on month-year constants, and memory beads decay along an Ebbinghaus curve instead of being truncated.

Q4：信任上限是什么？Q4: What is the trust ceiling?

对交互对象信任度的硬上限；经历创伤后上限下降，即使当下表现良好，信任也难以快速回到原来水平——更接近真实的关系动力学。The hard cap on how much trust an agent can hold for a counterpart. After trauma the ceiling drops, so trust cannot quickly return to previous levels even with good behavior — closer to real relational dynamics.

Q5：怎么接入自己的产品或游戏？Q5: How do I integrate it into my product / game?

纯 JS 引擎，3 行代码接入；详见工作台「架构」与「API」章节。支持任意运行时嵌入。Pure JS — 3 lines to integrate; see the Architecture and API sections of the workbench. Embeds in any runtime.

6. 术语表6. Glossary