necigit/Coral-Memory

necigit★ 0PythonLast synced: 2026-08-16

Open on GitHub

Heat-aware memory layer for LLM agents: three-tier hot/warm/cold storage, fusion retrieval (vector cosine + keyword Jaccard + time decay), heat-based eviction with an LLM-distill hook, hot-reload config. Ships as a DeepSeek Harness cordis plugin. MIT, single file, numpy-only.

README excerpt

脑珊瑚 · Coral Memory (Brain Coral) A heat-aware persistent memory layer for LLM agents — 面向 LLM Agent 的记忆层中间件: 三级存储(热/温/冷)、多路融合检索、热度生命周期淘汰、配置热加载、推理线索链路(永不遗忘的跨聊天协作)、DSH Harness 插件集成。 Origin: started as a ComfyUI prompt-manager idea, grew into a memory layer. 本来只想管提示词,结果长成了一片珊瑚礁。 EN : Coral Memory is a heat-aware persistent memory layer for LLM agents — three-tier storage (hot/warm/cold), multi-fusion retrieval, heat-based lifecycle eviction, config hot-reload, and a zero-dependency MCP stdio bridge so any MCP-capable client (DSH Harness, Claude Desktop, Cline, ...) gets memory search / memory insert tools. Embedding models are fetched on first run from HuggingFace — never committed to the repo. Author — Mr. Code Muggle (@Ne) Mr. Code Muggle — hi guys, I made something fun to play with: fork it, break it, rebuild it — just maybe mention me (lol). The coral remembers what I can't. Questions? 📮 [email protected] Shoutout to every open-source maker out there 🌱 这是什么 / 不是什么 是什么 :给 LLM Agent 用的"记忆层"——记住该记住的(你的偏好、说过的话), 忘掉该忘掉的(没用的旧记忆,按热度淘汰),并在需要时只捞最相关的几条给你。 不占上下文窗口,跨会话不丢失。 不是什么 : - 不是 RAG 框架(不负责分块、文档摄取、生成); - 不是向量数据库(无 ANN 索引,单机 20 万条以内的内存全量打分); - 不是缓存插件(缓存是"别重复计算",它是"别重复交代"); - 它不产生答案,它只负责"…

View full README on GitHub →
Agentsdsh-plugindsh-pluginsagentsearch

Category