基于 LLM、记忆系统与反思机制的成长型智能体
A self-growth agent powered by LLM, memory, and reflection.
中文
InnerTrace 是一个面向自我成长场景的 AI 系统，帮助你将日常记录转化为结构化认知，并通过长期记忆与证据检索发现行为模式、情绪波动与潜在改进方向。
它不只是“记录工具”，而是一个会基于你的历史数据进行推理的反思助手。
English
InnerTrace is an AI-powered self-growth system that turns daily journaling into structured understanding.
With long-term memory and evidence retrieval, it helps identify patterns, emotional shifts, and actionable improvements.
It is not just a note-taking tool, but a reflection assistant that reasons over your own historical data.
- 后端仓库 / Backend: https://github.com/yaruyng/InnerTrace
- 前端仓库 / Frontend: https://github.com/yaruyng/InnerTrace_frontend
- 自由记录每天经历，无固定模板
Write freely about your day with no strict format. - 自动抽取情绪、事件、主题、压力/能量等信号
Automatically extract signals such as emotion, events, topics, stress, and energy.
- 多层记忆：原始日志、结构化洞察、周期性总结
Multi-layer memory: raw entries, structured insights, and periodic summaries. - 当前已支持基于结构化标签与时间范围的检索；向量语义检索正在建设中
Currently supports structured tag + time-range retrieval; vector semantic retrieval is in progress.
- 针对“为什么最近疲惫”“什么在消耗精力”等问题进行证据归因
Provides evidence-grounded answers to questions like "Why have I felt tired recently?" - 输出模式分析与可执行建议
Returns pattern analysis and actionable suggestions.
- 汇总情绪趋势、重复问题、亮点与风险
Summarizes emotional trends, repeated issues, highlights, and risks.
Journal Input
↓
LLM Structured Analysis
↓
Memory Storage (MySQL + Vector DB)
↓
Structured Retrieval (Vector Retrieval In Progress)
↓
Agent Reasoning
↓
Insight + Evidence + Suggestions
User Question
↓
Intent Recognition
↓
Memory Retrieval (Vector + Metadata)
Memory Retrieval (Metadata/Structured First, Vector In Progress)
↓
Evidence Aggregation
↓
LLM Reasoning
↓
Structured Response
设计目标 / Goals:
- 按需检索记忆，而非盲目生成
Retrieve memory only when needed. - 选择有效时间窗口与主题上下文
Select relevant time windows and topic context. - 确保结论有历史证据支撑
Keep conclusions grounded in historical evidence.
- Java 17
- Spring Boot
- MyBatis-Plus
- MySQL (结构化数据 / structured data)
- pgvector / Milvus (向量记忆，建设中 / vector memory, in progress)
- Redis (缓存 / cache)
- LLM API (分析与推理 / analysis and reasoning)
- Embedding model
- 异步任务处理 / Async task processing
下图为 InnerTrace Web 界面示意。
Frontend
↓
API Layer
↓
InnerTrace Agent Service
├─ Journal Module
├─ Analysis Module
├─ Memory Module
├─ Reflection Module
└─ Agent Module
↓
Storage Layer
├─ MySQL
├─ Vector DB
└─ Redis
-
基于数据，而非臆测 | No hallucinated psychology
只依据用户历史记录进行反思，不凭空贴标签。
Reflect only from user data, not assumptions. -
证据优先 | Evidence-first insight
每个关键结论应可追溯到具体记录。
Key conclusions should be traceable to concrete records. -
温和引导 | Gentle guidance, not judgment
提供可执行建议，而非价值评判。
Provide actionable guidance rather than judgment.
- 日志输入与存储 / Journal input & storage
- 结构化分析 / Structured analysis
- 向量记忆检索 / Vector memory retrieval
- 周/月度反思 / Weekly & monthly reflection
- 多智能体推理 / Multi-agent reasoning
- 个性化调优 / Personalization tuning
- 可视化看板 / Visualization dashboard
中文
-
准备后端（端口 8080）
- 安装 Java 17 与 Maven
- 在后端目录执行：
cd InnerTrace mvn clean package java -jar target/*.jar
- 或直接开发模式：
mvn spring-boot:run
- 请根据本地环境修改
application.yml中的数据库与 LLM API Key 配置。
-
启动前端（端口 3000）
- 安装 Node.js（推荐 18+）
- 在前端目录执行：
cd innertrace_frontend npm install npm run dev - 打开浏览器访问
http://localhost:3000
English
-
Backend (port 8080)
- Install Java 17 and Maven.
- From the backend directory:
cd InnerTrace mvn clean package java -jar target/*.jar
- Or run in dev mode:
mvn spring-boot:run
- Configure your own database and LLM API key in
application.ymlbefore running.
-
Frontend (port 3000)
- Install Node.js (18+ recommended).
- From the frontend directory:
cd innertrace_frontend npm install npm run dev - Visit
http://localhost:3000in your browser.
欢迎提交 Issue 和 PR，一起完善 AI 驱动的长期反思系统。
Issues and pull requests are welcome to improve this AI-driven long-term reflection system.
Apache-2.0