AgentSkillsCN

agent-memory

实现智能体内存——短期、长期、语义存储与检索

SKILL.md
--- frontmatter
name: agent-memory
description: Implement agent memory - short-term, long-term, semantic storage, and retrieval
sasmp_version: "1.3.0"
bonded_agent: 06-agent-memory
bond_type: PRIMARY_BOND
version: "2.0.0"

Agent Memory

Give agents the ability to remember and learn across conversations.

When to Use This Skill

Invoke this skill when:

  • Adding conversation history
  • Implementing long-term memory
  • Building personalized agents
  • Managing context windows

Parameter Schema

ParameterTypeRequiredDescriptionDefault
taskstringYesMemory goal-
memory_typeenumNobuffer, summary, vector, hybridhybrid
persistenceenumNosession, user, globalsession

Quick Start

python
from langchain.memory import ConversationBufferWindowMemory

# Simple buffer (last k messages)
memory = ConversationBufferWindowMemory(k=10)

# With summarization
from langchain.memory import ConversationSummaryBufferMemory
memory = ConversationSummaryBufferMemory(llm=llm, max_token_limit=2000)

# Vector store memory
from langchain.memory import VectorStoreRetrieverMemory
memory = VectorStoreRetrieverMemory(retriever=vectorstore.as_retriever())

Memory Types

TypeUse CaseProsCons
BufferShort chatsSimpleNo compression
SummaryLong chatsCompactLoses detail
VectorSemantic recallRelevantSlower
HybridProductionBest of allComplex

Multi-Layer Architecture

python
class ProductionMemory:
    def __init__(self):
        self.short_term = BufferMemory(k=10)    # Recent
        self.summary = SummaryMemory()           # Compressed
        self.long_term = VectorMemory()          # Semantic

Troubleshooting

IssueSolution
Context overflowAdd summarization
Slow retrievalCache, reduce k
Irrelevant recallImprove embeddings
Memory not persistingCheck storage backend

Best Practices

  • Use multi-layer memory for production
  • Set token limits to prevent overflow
  • Add metadata (timestamps, importance)
  • Implement TTL for old memories

Related Skills

  • rag-systems - Vector retrieval
  • llm-integration - Context management
  • ai-agent-basics - Agent architecture

References