记忆

用于 AI 代理的长期语义记忆，使用 adk-memory。

概述

记忆系统为代理对话提供持久的、可搜索的存储。与会话状态（它是短暂的）不同，记忆在会话之间持久存在，并使代理能够从过去的交互中回忆起相关的上下文。

安装

[dependencies]
adk-memory = "2.0.0"

核心概念

MemoryEntry

一个包含内容、作者和时间戳的单一记忆记录：

use adk_memory::MemoryEntry;
use adk_core::Content;
use chrono::Utc;

let entry = MemoryEntry {
    content: Content::new("user").with_text("I prefer dark mode"),
    author: "user".to_string(),
    timestamp: Utc::now(),
};

MemoryService Trait

记忆后端的核⼼ Trait：

#[async_trait]
pub trait MemoryService: Send + Sync {
    /// Store session memories for a user
    async fn add_session(
        &self,
        app_name: &str,
        user_id: &str,
        session_id: &str,
        entries: Vec<MemoryEntry>,
    ) -> Result<()>;

    /// Search memories by query
    async fn search(&self, req: SearchRequest) -> Result<SearchResponse>;
}

SearchRequest

记忆搜索的查询参数：

use adk_memory::SearchRequest;

let request = SearchRequest {
    query: "user preferences".to_string(),
    user_id: "user-123".to_string(),
    app_name: "my_app".to_string(),
    limit: None,
    min_score: None,
    project_id: None, // None = global only, Some("id") = global + project
};

InMemoryMemoryService

用于开发和测试的简单内存实现：

use adk_memory::{InMemoryMemoryService, MemoryService, MemoryEntry, SearchRequest};
use adk_core::Content;
use chrono::Utc;

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let memory = InMemoryMemoryService::new();

    // Store memories from a session
    let entries = vec![
        MemoryEntry {
            content: Content::new("user").with_text("I like Rust programming"),
            author: "user".to_string(),
            timestamp: Utc::now(),
        },
        MemoryEntry {
            content: Content::new("assistant").with_text("Rust is great for systems programming"),
            author: "assistant".to_string(),
            timestamp: Utc::now(),
        },
    ];

    memory.add_session("my_app", "user-123", "session-1", entries).await?;

    // Search memories
    let request = SearchRequest {
        query: "Rust".to_string(),
        user_id: "user-123".to_string(),
        app_name: "my_app".to_string(),
        limit: None,
        min_score: None,
        project_id: None,
    };

    let response = memory.search(request).await?;
    println!("Found {} memories", response.memories.len());

    Ok(())
}

记忆隔离

记忆通过以下方式隔离：

app_name：不同的应用程序拥有独立的记忆空间
user_id：每个用户的记忆都是私有的
project_id（可选）：条目可以限定在用户内的某个项目

// User A's memories
memory.add_session("app", "user-a", "sess-1", entries_a).await?;

// User B's memories (separate)
memory.add_session("app", "user-b", "sess-1", entries_b).await?;

// Search only returns user-a's memories
let request = SearchRequest {
    query: "topic".to_string(),
    user_id: "user-a".to_string(),
    app_name: "app".to_string(),
    limit: None,
    min_score: None,
    project_id: None, // None = global entries only
};

项目范围记忆

记忆可以限定在用户内的某个项目。隔离键变为 (app_name, user_id, project_id?)：

全局条目 (project_id = None)：在所有项目上下文和仅全局搜索中可见。
项目条目 (project_id = Some(id))：仅在该特定项目内搜索时可见。
项目搜索 (project_id = Some(id))：返回全局条目 + 该项目的条目。
全局搜索 (project_id = None)：仅返回全局条目。

存储项目范围条目

use adk_memory::{InMemoryMemoryService, MemoryService, MemoryEntry};
use adk_core::Content;
use chrono::Utc;

let service = InMemoryMemoryService::new();

let entry = MemoryEntry {
    content: Content::new("user").with_text("Project uses microservices"),
    author: "user".to_string(),
    timestamp: Utc::now(),
};

// Global entry (no project scope)
service.add_session("app", "user-1", "sess-1", vec![entry.clone()]).await?;

// Project-scoped entry
service.add_session_to_project("app", "user-1", "sess-2", "my-project", vec![entry.clone()]).await?;

// Single entry to a project
service.add_entry_to_project("app", "user-1", "my-project", entry).await?;

使用项目范围进行搜索

use adk_memory::SearchRequest;

// Global-only search — returns only global entries
let global = service.search(SearchRequest {
    query: "microservices".into(),
    user_id: "user-1".into(),
    app_name: "app".into(),
    limit: None,
    min_score: None,
    project_id: None,
}).await?;

// Project search — returns global + project entries
let project = service.search(SearchRequest {
    query: "microservices".into(),
    user_id: "user-1".into(),
    app_name: "app".into(),
    limit: None,
    min_score: None,
    project_id: Some("my-project".into()),
}).await?;

项目范围删除

// Delete entries matching a query within a project only
service.delete_entries_in_project("app", "user-1", "my-project", "microservices").await?;

// Delete ALL entries for a project
service.delete_project("app", "user-1", "my-project").await?;

// Global delete — only removes global entries, project entries are unaffected
service.delete_entries("app", "user-1", "microservices").await?;

// GDPR delete_user — removes everything (global + all projects)
service.delete_user("app", "user-1").await?;

MemoryServiceAdapter 与项目范围

MemoryServiceAdapter 将 MemoryService 连接到 adk_core::Memory。使用 with_project_id() 来限定所有操作的范围：

use adk_memory::{InMemoryMemoryService, MemoryServiceAdapter};
use adk_core::Memory;
use std::sync::Arc;

let service = Arc::new(InMemoryMemoryService::new());

// Adapter without project — operates on global entries
let global_adapter = MemoryServiceAdapter::new(service.clone(), "app", "user-1");

// Adapter with project — all search/add/delete operations scoped to the project
let project_adapter = MemoryServiceAdapter::new(service.clone(), "app", "user-1")
    .with_project_id("my-project");

// Core Memory trait also supports ad-hoc project access
global_adapter.search_in_project("query", "other-project").await?;
global_adapter.add_to_project(entry, "other-project").await?;

项目 ID 验证

项目标识符在所有写入操作中都会进行验证：

不得为空
不得超过 256 个字符

use adk_memory::validate_project_id;

validate_project_id("my-project")?;       // Ok
validate_project_id("")?;                  // Err: must not be empty
validate_project_id(&"x".repeat(257))?;   // Err: exceeds 256 chars

搜索语义矩阵

`SearchRequest.project_id`	返回全局条目	返回项目条目
`None`	✅ 匹配查询	❌ 无
`Some("A")`	✅ 匹配查询	✅ 仅项目“A”中匹配查询的条目

删除语义矩阵

操作	范围
`delete_entries` (无项目)	仅匹配查询的全局条目
`delete_entries_in_project("A")`	仅匹配查询的项目 "A" 条目
`delete_project("A")`	项目 "A" 的所有条目
`delete_user`	所有条目（全局 + 所有项目）

搜索行为

InMemoryMemoryService 使用基于词语的匹配：

查询被分词为单词（小写）
每个记忆的内容被分词
返回包含任何匹配单词的记忆

// Query: "rust programming"
// Matches memories containing "rust" OR "programming"

自定义记忆后端

实现 MemoryService 以用于自定义存储（例如，向量数据库）：

use adk_memory::{MemoryService, MemoryEntry, SearchRequest, SearchResponse};
use adk_core::Result;
use async_trait::async_trait;

pub struct VectorMemoryService {
    // Your vector DB client
}

#[async_trait]
impl MemoryService for VectorMemoryService {
    async fn add_session(
        &self,
        app_name: &str,
        user_id: &str,
        session_id: &str,
        entries: Vec<MemoryEntry>,
    ) -> Result<()> {
        // 1. Generate embeddings for each entry
        // 2. Store in vector database with metadata
        Ok(())
    }

    async fn search(&self, req: SearchRequest) -> Result<SearchResponse> {
        // 1. Generate embedding for query
        // 2. Perform similarity search
        // 3. Return top-k results
        Ok(SearchResponse { memories: vec![] })
    }
}

与 Agent 的集成

记忆与 LlmAgentBuilder 集成：

use adk_agent::LlmAgentBuilder;
use adk_memory::InMemoryMemoryService;
use std::sync::Arc;

let memory = Arc::new(InMemoryMemoryService::new());

let agent = LlmAgentBuilder::new("assistant")
    .model(model)
    .instruction("You are a helpful assistant with memory.")
    .memory(memory)
    .build()?;

当记忆被配置时：

在每个回合之前，搜索相关记忆
匹配的记忆被注入到上下文中
在每个会话之后，对话被存储为记忆

架构

┌─────────────────────────────────────────────────────────────┐
│                      Agent Request                          │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                    Memory Search                            │
│                                                             │
│   SearchRequest { query, user_id, app_name, project_id }   │
│                         │                                   │
│                         ▼                                   │
│   ┌─────────────────────────────────────────────────────┐  │
│   │              MemoryService                          │  │
│   │  ┌─────────┐ ┌──────────┐ ┌────────┐ ┌──────────┐  │  │
│   │  │InMemory │ │ SQLite   │ │Postgres│ │  Redis   │  │  │
│   │  │(dev)    │ │          │ │pgvector│ │          │  │  │
│   │  └─────────┘ └──────────┘ └────────┘ └──────────┘  │  │
│   │  ┌─────────┐ ┌──────────┐                           │  │
│   │  │MongoDB  │ │  Neo4j   │                           │  │
│   │  └─────────┘ └──────────┘                           │  │
│   └─────────────────────────────────────────────────────┘  │
│                         │                                   │
│                         ▼                                   │
│   SearchResponse { memories: Vec<MemoryEntry> }            │
│   (filtered by project scope)                              │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│              Context Injection                              │
│                                                             │
│   Relevant memories added to agent context                 │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                    Agent Execution                          │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                   Memory Storage                            │
│                                                             │
│   Session conversation stored for future recall            │
│   (global or project-scoped)                               │
└─────────────────────────────────────────────────────────────┘

最佳实践

实践	描述
在生产环境中使用向量数据库	InMemory 仅用于开发/测试
按用户范围划分	始终包含 user_id 以保护隐私
限制结果	限制返回的记忆以避免上下文溢出
清理旧记忆	对陈旧数据实施 TTL 或归档
策略性地嵌入	存储摘要，而非原始对话

与会话的比较

特性	会话状态	记忆
持久性	会话生命周期	永久
范围	单个会话	跨会话
搜索	键值查找	语义搜索
用例	当前上下文	长期回忆

上一页: ← Guardrails | 下一页: Studio →