Persistent Memory | HyperKit.ai

Definition

Persistent memory refers to information stored outside the immediate context window that survives across sessions. Unlike working memory that clears with each conversation, persistent memory allows the model to recall user preferences, past interactions, and learned facts over time.

This is typically implemented through external databases, embeddings, or explicit memory storage systems.

Key Concepts

Cross-session retention: Information persists between conversations
External storage: Database, vector store, or file system
Retrieval mechanism: How memories are recalled when relevant
Memory updates: Adding, modifying, or forgetting information

Examples

Architecture

Persistent Memory System

Session 1 (Monday):
User: "I prefer Python over JavaScript"
System: [Stores preference in memory DB]

Session 2 (Wednesday):
User: "What language should I use for this script?"
System: [Retrieves: "User prefers Python"]
Assistant: "Given your preference for Python, I'd suggest..."

Memory Database Structure:
┌─────────────────────────────────────────┐
│ user_id: "alice_123"                    │
│ memories: [                             │
│   {                                     │
│     type: "preference",                 │
│     content: "Prefers Python over JS",  │
│     timestamp: "2024-01-15",            │
│     confidence: 0.95                    │
│   },                                    │
│   {                                     │
│     type: "fact",                       │
│     content: "Works at TechCorp",       │
│     timestamp: "2024-01-10",            │
│     confidence: 0.90                    │
│   }                                     │
│ ]                                       │
└─────────────────────────────────────────┘

Implementation

Memory Management

class PersistentMemory:
    def __init__(self, user_id, vector_db):
        self.user_id = user_id
        self.db = vector_db

    def store(self, content, memory_type="general"):
        """Save a new memory"""
        embedding = embed(content)
        self.db.insert({
            "user_id": self.user_id,
            "content": content,
            "type": memory_type,
            "embedding": embedding,
            "timestamp": now()
        })

    def recall(self, query, k=5):
        """Retrieve relevant memories"""
        query_embedding = embed(query)
        return self.db.search(
            query_embedding,
            filter={"user_id": self.user_id},
            limit=k
        )

    def inject_into_prompt(self, query, system_prompt):
        """Add memories to working memory"""
        memories = self.recall(query)
        memory_context = "\n".join([m.content for m in memories])
        return f"{system_prompt}\n\nUser context:\n{memory_context}"

Interactive Exercise

✎

Design Memory Categories

For a personal assistant, what categories of information should be stored in persistent memory?

Pro Tips

Distinguish facts (stable) from preferences (may change)
Include timestamps for temporal relevance
Implement confidence scores for uncertain memories
Allow users to view and delete their stored memories

Definition

Key Concepts

Examples

Interactive Exercise

Related Terms