fzs356113-oss/dsh-self-evolving-agent

README 摘要

dsh-self-evolving-agent Self-Evolving Agent · 自我进化 Agent — 把「自我改进 Agent」模式做成 DSH 原生框架: 三本日志 + 六触发场景 + 经验晋升管线 + soul.md 身份层 + 选择性经验注入。 English overview: This repository is a plugin for DeepSeek Harness (DSH), the 122k+ star agent framework where everything is a plugin . It turns a stateless DSH agent into a self-evolving agent : without touching model weights, it records failures, corrections and best practices into plain-text memory files (three logs), selectively injects only the experience relevant to the current task (never full-context dumps), and promotes repeatedly-verified lessons up a pipeline — from raw candidates ( .learnings/ ) to project rules (AGENTS.md), identity behaviors (SOUL.md) and finally reusable skills (skills/). It ships two ready-made DSH workflows ( learn-from-session , promotion-review with dry-run safety) plus an orchestrating skill ( self-evolving-loop ). Everything is UTF-8 Markdown/JSON — human readable and git diffable. ✨ 亮点 1. 不改权重,靠文件记忆 — Agent 最大的结构性缺陷是"无状态": 每个会话从零开始,重复踩同一个坑。本框架用文件系统做记忆: 错误自动记录 → 模式识别 → 晋升为永久规则 → 提取为可复用技能,越用越聪明。 2. 四层记忆 + 选择性注入 — 会话工作记忆(HOT)/ 三本日志(WARM)/ 行为规则 / 可复用技能(COLD)分层管理;每次只 grep 与当前任务相关的经验注入上下文,绝不全量注入(呼应「智能RAG选择性检索」: 精准而非全面…

分类