Memory

Categories

• Knowledge Encoded in Model Parameters
  • Knowledge acquired during training and stored implicitly in model weights.
• Short-Term Memory
  • Information currently available in the model’s context window.
• Long-Term Memory
  • Information persisted outside the model and retrieved when needed.

Why Is Memory Needed?

• Context windows have finite capacity.
• Agents need to preserve task history and intermediate state.
• Useful information should remain available across sessions.

What Is Memory?

• A mechanism for storing, retrieving, updating, and using information.

Memory Management

• Store and Update Information
  • Summarize and Remove Duplicates
  • Expire or Forget Outdated Information
  • Resolve Conflicting Memories

Memory Retrieval

• Database Queries or Search
• Retrieval-Augmented Search