memory-systems

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Memory System Design

内存系统设计

Memory provides persistence that allows agents to maintain continuity across sessions and reason over accumulated knowledge.

内存提供持久化能力，让Agent能够在跨会话场景下保持连续性，并基于累积的知识进行推理。

Memory Architecture Spectrum

内存架构范围

Layer	Latency	Persistence	Use Case
Working Memory	Zero	Volatile	Context window
Short-Term	Low	Session	Session state
Long-Term	Medium	Persistent	Cross-session knowledge
Entity Memory	Medium	Persistent	Entity tracking
Temporal KG	Medium	Persistent	Time-aware queries

层级	延迟	持久化能力	适用场景
工作内存	零延迟	易失性	上下文窗口
短期内存	低延迟	会话级	会话状态
长期内存	中等延迟	持久化	跨会话知识
实体内存	中等延迟	持久化	实体跟踪
时序知识图谱（Temporal KG）	中等延迟	持久化	时间感知查询

Memory System Performance

内存系统性能

System	DMR Accuracy	Retrieval Latency
Zep (Temporal KG)	94.8%	2.58s
MemGPT	93.4%	Variable
GraphRAG	75-85%	Variable
Vector RAG	60-70%	Fast
Recursive Summary	35.3%	Low

系统	DMR准确率	检索延迟
Zep（时序知识图谱）	94.8%	2.58秒
MemGPT	93.4%	可变
GraphRAG	75-85%	可变
向量RAG	60-70%	快速
递归摘要	35.3%	低

Why Vector Stores Fall Short

向量存储的局限性

Vector stores lose relationship information:

Can retrieve "Customer X purchased Product Y"
Cannot answer "What did customers who bought Y also buy?"
Cannot distinguish current vs outdated facts

向量存储会丢失关系信息：

可以检索到「客户X购买了产品Y」
无法回答「购买Y的客户还买了什么？」
无法区分当前事实与过时事实

Memory Implementation Patterns

内存实现模式

Pattern 1: File-System-as-Memory

模式1：文件系统作为内存

python

undefined

python

undefined

Simple, no infrastructure needed

def store_fact(entity_id, fact): path = f"memory/{entity_id}.json" facts = load_json(path, default=[]) facts.append({"fact": fact, "timestamp": now()}) save_json(path, facts)

undefined

def store_fact(entity_id, fact): path = f"memory/{entity_id}.json" facts = load_json(path, default=[]) facts.append({"fact": fact, "timestamp": now()}) save_json(path, facts)

undefined

Pattern 2: Vector RAG with Metadata

模式2：带元数据的向量RAG

python

undefined

python

undefined

Embed facts with rich metadata

vector_store.add( embedding=embed(fact), metadata={ "entity_id": entity_id, "valid_from": now(), "source": "conversation", "confidence": 0.95 } )

undefined

vector_store.add( embedding=embed(fact), metadata={ "entity_id": entity_id, "valid_from": now(), "source": "conversation", "confidence": 0.95 } )

undefined

Pattern 3: Knowledge Graph

模式3：知识图谱

python

undefined

python

undefined

Preserve relationships

graph.create_relationship( from_entity="Customer_123", relationship="PURCHASED", to_entity="Product_456", properties={"date": "2024-01-15", "quantity": 2} )

undefined

graph.create_relationship( from_entity="Customer_123", relationship="PURCHASED", to_entity="Product_456", properties={"date": "2024-01-15", "quantity": 2} )

undefined

Pattern 4: Temporal Knowledge Graph

模式4：时序知识图谱

python

undefined

python

undefined

Time-travel queries

def query_address_at_time(user_id, query_time): return graph.query(""" MATCH (user)-[r:LIVES_AT]->(address) WHERE user.id = $user_id AND r.valid_from <= $query_time AND (r.valid_until IS NULL OR r.valid_until > $query_time) RETURN address """, {"user_id": user_id, "query_time": query_time})

undefined

undefined

Entity Memory

实体内存

Track entities consistently across conversations:

Entity Identity: "John Doe" in one conversation = same person in another
Entity Properties: Facts discovered about entities over time
Entity Relationships: Relationships discovered between entities

python

def remember_entity(entity_id, properties):
    memory.store({
        "type": "entity",
        "id": entity_id,
        "properties": properties,
        "last_updated": now()
    })

跨会话一致跟踪实体：

实体身份：一次对话中的「John Doe」等同于另一次对话中的同一个人
实体属性：随时间推移发现的实体相关事实
实体关系：发现的实体之间的关系

python

def remember_entity(entity_id, properties):
    memory.store({
        "type": "entity",
        "id": entity_id,
        "properties": properties,
        "last_updated": now()
    })

Memory Consolidation

内存整合

Trigger consolidation when:

Memory accumulates significantly
Retrieval returns too many outdated results
Periodically on schedule
Explicit request

Process:

Identify outdated facts
Merge related facts
Update validity periods
Archive/delete obsolete facts
Rebuild indexes

在以下场景触发整合：

内存累积量显著增加时
检索返回过多过时结果时
按计划定期执行
收到明确请求时

流程：

识别过时事实
合并相关事实
更新有效期
归档/删除过时事实
重建索引

Choosing Memory Architecture

选择内存架构

Requirement	Architecture
Simple persistence	File-system memory
Semantic search	Vector RAG with metadata
Relationship reasoning	Knowledge graph
Temporal validity	Temporal knowledge graph

需求	架构
简单持久化	文件系统内存
语义搜索	带元数据的向量RAG
关系推理	知识图谱
时序有效性	时序知识图谱

Best Practices

最佳实践

Match architecture to query requirements
Implement progressive disclosure for access
Use temporal validity to prevent conflicts
Consolidate periodically
Design for retrieval failures gracefully
Consider privacy implications
Implement backup and recovery
Monitor growth and performance

根据查询需求匹配架构
实现渐进式披露以控制访问
使用时序有效性避免冲突
定期执行整合操作
优雅设计检索失败的处理逻辑
考虑隐私影响
实现备份与恢复机制
监控内存增长与性能