解决 Agent 记忆问题的魔杖
2 分•作者: fokkedekker•5 个月前
我与数百位 AI 智能体开发者交流过，对于“如果给你一根魔杖解决一个问题，你会选择什么？”这个问题，他们的答案都是智能体记忆。<p>我们在 Raindrop 中构建了 SmartMemory 来解决这个问题，它为智能体提供了四种协同工作的记忆类型：<p>记忆类型概述<p>工作记忆
• 在会话过程中保存活跃的上下文
• 将想法组织成不同的时间线（主题）
• 智能体可以搜索你讨论过的内容，并在之前的观点基础上进行构建
• 就像正在进行的对话的短期记忆<p>情景记忆
• 将已完成的会话存储为可搜索的历史记录
• 记住几周或几个月前你讨论过的内容
• 可以恢复之前的对话，以便在中断的地方继续
• 你的智能体的长期对话档案<p>语义记忆
• 存储事实、文档和参考资料
• 在所有对话中保持知识
• 积累关于你的项目和偏好的信息
• 你的智能体的知识库，随着时间的推移而增长<p>程序记忆
• 保存工作流程、工具交互模式和程序
• 学会如何一致地处理不同的情况
• 存储决策树和响应模式
• 你的智能体学习的技能和操作程序<p>真正有效的多层搜索<p>*工作记忆* 使用嵌入和向量搜索。当你搜索“身份验证问题”时，它会找到关于“登录问题”或“安全漏洞”的记忆，即使确切的单词不匹配。<p>*情景记忆、语义记忆和程序记忆* 使用三层搜索方法：
• 基于语义含义的向量搜索
• 基于提取的实体和关系的图搜索
• 关键词和主题匹配，用于精确查询<p>这种多层方法意味着你的智能体可以找到相关信息，无论你是通过概念、想法之间的特定关系还是确切的术语进行搜索。<p>使用 SmartMemory 的三种方式<p>选项 1：完整的 Raindrop 框架
在 Raindrop 中构建你的智能体，并获得完整的记忆系统以及其他智能体基础设施：<p>```hcl
application "my-agent" {
smartmemory "agent_memory" {}
}<p>```<p>选项 2：MCP 集成
已经有一个智能体了？将我们的 MCP（模型上下文协议）服务器连接到你现有的设置。启动一个 SmartMemory 实例，你的智能体就可以通过 MCP 调用访问所有记忆功能——无需重建任何东西。<p>选项 3：API/SDK 如果你已经有一个智能体，但对 MCP 不熟悉，我们也有一个简单的 API 和 SDK（Python、TypeScript、Java 和 Go）供你使用<p>一些有用的链接，帮助你开始<p>注册请访问：https://liquidmetal.ai/
概念文档请访问：https://docs.liquidmetal.ai/concepts/smartmemory/
实现文档请访问：https://docs.liquidmetal.ai/reference/resources/smartmemory/
快速入门请访问：https://docs.liquidmetal.ai/tutorials/smartmemory-app-deployment/
查看原文
I spoke to hundreds of AI agent developers and the answer to the question - "if you had one magic wand to solve one thing, what would it be?" - was agent memory.<p>We built SmartMemory in Raindrop to solve this problem by giving agents four types of memory that work together:<p>We built SmartMemory in Raindrop to solve this problem by giving agents four types of memory that work together:<p>Memory Types Overview<p>Working Memory
• Holds active conversation context within sessions
• Organizes thoughts into different timelines (topics)
• Agents can search what you've discussed and build on previous points
• Like short-term memory for ongoing conversations<p>Episodic Memory
• Stores completed conversation sessions as searchable history
• Remembers what you discussed weeks or months ago
• Can restore previous conversations to continue where you left off
• Your agent's long-term conversation archive<p>Semantic Memory
• Stores facts, documents, and reference materials
• Persists knowledge across all conversations
• Builds up information about your projects and preferences
• Your agent's knowledge base that grows over time<p>Procedural Memory
• Saves workflows, tool interaction patterns, and procedures
• Learns how to handle different situations consistently
• Stores decision trees and response patterns
• Your agent's learned skills and operational procedures<p>Multi-Layer Search That Actually Works<p>*Working Memory* uses embeddings and vector search. When you search for "authentication issues," it finds memories about "login problems" or "security bugs" even though the exact words don't match.<p>*Episodic, Semantic, and Procedural Memory* use a three-layer search approach:
• Vector search for semantic meaning
• Graph search based on extracted entities and relationships
• Keyword and topic matching for precise queries<p>This multi-layer approach means your agent can find relevant information whether you're searching by concept, by specific relationships between ideas, or by exact terms.<p>Three Ways to Use SmartMemory<p>Option 1: Full Raindrop Framework
Build your agent within Raindrop and get the complete memory system plus other agent infrastructure:<p>```hcl
application "my-agent" {
smartmemory "agent_memory" {}
}<p>```<p>Option 2: MCP Integration
Already have an agent? Connect our MCP (Model Context Protocol) server to your existing setup. Spin up a SmartMemory instance and your agent can access all memory functions through MCP calls - no need to rebuild anything.<p>Option 3: API/SDK If you already have an agent but are not familar with MCP we also have a simple API and SDK (pytyon, TypeScript, Java and Go) you can use<p>A couple of helpful links to get started<p>For signup check: https://liquidmetal.ai/
For concepts documentation check: https://docs.liquidmetal.ai/concepts/smartmemory/
For implementation documentation check: https://docs.liquidmetal.ai/reference/resources/smartmemory/
For quick start check https://docs.liquidmetal.ai/tutorials/smartmemory-app-deployment/