Multi-tiered Memory Framework

Multi-tiered Memory Architecture for LLM Agents

Definition

An async memory management system that organizes agent memory across multiple tiers or layers, enabling efficient storage, retrieval, and context management for long-running LLM agent workflows. Sawtooth exemplifies this pattern by providing structured memory hierarchies that reduce latency and improve agent reasoning over extended interactions.

Examples in the Wild

  • Example 1:Sawtooth's async multi-tiered memory for managing conversation history, tool outputs, and reasoning traces
  • Example 2:Tiered memory systems that separate hot (frequently accessed) from cold (archived) agent state
  • Example 3:Memory frameworks that optimize token usage by intelligently promoting/demoting context across tiers