NanoMemAgent — modular agent framework with layered memory
Python · ReAct · Skill / Tool / MCP · ChromaDB · Neo4j · BGE-M3 · BM25 · ThreadPool · 2026.02 – 2026.03
A modular agent execution framework built from scratch: a ReAct reasoning loop, dynamic tool calls, MCP integration, a layered memory system (short-term vector store + long-term knowledge graph), and a plug-in Skill mechanism. Supports multi-model switching; already wired into Feishu.
Agent Loop engine. The core module coordinates message routing → ContextBuilder (system prompt + memory + skills) → LLM inference → tool-call execution → feedback, with max-iteration guards and chained tool calls.
Layered memory (L1–L5). Storage / embedding / retrieval / knowledge / application. Dual-track storage: Operational Memory (dynamic CRUD) + RAG Knowledge Base. Retrieval runs vector + BM25 + graph in parallel, fused with RRF ranking.
Tool ecosystem. A standardized MCP Client lets filesystem / DB / search tools plug in with no framework changes. Built-ins: cron scheduling, background subagents, heartbeat health checks. A single LLMProvider interface swaps OpenAI / Anthropic / local models with zero call-site changes.
从零实现的模块化 Agent 执行框架。内置 ReAct 推理循环、动态 tool 调用与 MCP 接入、分层记忆系统（短期向量存储 + 长期知识图谱）、插件化 Skill 机制，支持多模型切换，并已接入飞书。
Agent Loop 核心引擎。 协调 消息路由 → ContextBuilder（系统提示 + 记忆 + 技能）→ LLM 推理 → Tool Calls 执行 → 结果反馈 的完整循环，支持最大迭代次数保护与工具链式调用。
分层记忆系统（L1–L5）。 存储 / 嵌入 / 检索 / 知识 / 应用 五层架构；Operational Memory（动态 CRUD）与 RAG Knowledge Base 双轨存储；Vector + BM25 + Graph 三路并行检索 + RRF 融合排序。
模块化工具生态。 标准化 MCP Client 支持文件系统、数据库、搜索等外部工具动态接入；内置 Cron 定时任务、子代理（Subagent）后台执行、Heartbeat 健康检测；抽象 LLMProvider 接口层统一封装 OpenAI、Anthropic、本地模型等，零代码切换。