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
| Parameter | Type | Required | Description | Default |
|---|
task | string | Yes | Memory goal | - |
memory_type | enum | No | buffer, summary, vector, hybrid | hybrid |
persistence | enum | No | session, user, global | session |
Quick Start
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
| Type | Use Case | Pros | Cons |
|---|
| Buffer | Short chats | Simple | No compression |
| Summary | Long chats | Compact | Loses detail |
| Vector | Semantic recall | Relevant | Slower |
| Hybrid | Production | Best of all | Complex |
Multi-Layer Architecture
class ProductionMemory:
def __init__(self):
self.short_term = BufferMemory(k=10) # Recent
self.summary = SummaryMemory() # Compressed
self.long_term = VectorMemory() # Semantic
Troubleshooting
| Issue | Solution |
|---|
| Context overflow | Add summarization |
| Slow retrieval | Cache, reduce k |
| Irrelevant recall | Improve embeddings |
| Memory not persisting | Check 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 retrievalllm-integration - Context managementai-agent-basics - Agent architecture
References
