Agent Memory System
Give your AI agent persistent, long-term memory across conversations and sessions.
Memory Types Implemented
Episodic Memory
Stores episodes/events from conversations:
- Key facts extracted per conversation
- Decisions made and context
- User preferences and patterns
- "Remembering" past interactions
Semantic Memory
Structured knowledge storage:
- Entity definitions and relationships
- Facts about the world
- Domain knowledge base
- Learned procedures
Procedural Memory
Agent's own capabilities:
- Known skills and tools
- How to use different APIs
- Response patterns that worked
Architecture
User Input
↓
Short-term (current session context)
↓
Memory Retrieval → Top-k relevant memories (vector search)
↓
Context Injection → Combined prompt
↓
LLM Response
↓
Memory Storage → Extract new facts, update entities
Features
- Vector-based storage (ChromaDB or Pinecone)
- Entity extraction (spaCy NER)
- Conversation summarization (every N turns)
- Relevance scoring for retrieval
- Forgetting/summarization of old memories
Use Cases
- Personal AI assistant that remembers you
- Customer support agent with context
- Research agent with persistent knowledge
- Trading agent with market memory
- Personal CRM (remembering people and their context)
Technical Stack
- ChromaDB / Pinecone (vector store)
- spaCy (entity extraction)
- LangChain (memory abstractions)
- PostgreSQL (structured memory)
Pricing
| Type | Context Window | Price |
|---|---|---|
| Basic | 100K tokens | $100 |
| Pro | 1M tokens | $300 |
| Enterprise | Unlimited | $800 |
Built by Beta