memory-systems

You can persist agent state across sessions by implementing a layered memory architecture with working, short-term, and long-term memory layers. This enables agents to maintain continuity and reason over accumulated knowledge instead of starting stateless interactions.

A layered memory architecture separates agent state into working, short-term, long-term, entity, and temporal graph layers. You need it when agents must track entities and reason over accumulated knowledge in dynamic environments across multiple conversations.

Implement temporal and semantic retrieval by storing memories with timestamps and entity tags, then fetch relevant context using memory_context or retrieve_memories APIs to display accumulated knowledge alongside current task context.

Yes, you can track entities and reason over time by building temporal knowledge graphs within the memory system. This enables reliable entity tracking and time-aware reasoning across multiple agent interactions and sessions.

Inject accumulated memories into agent decisions by integrating a context loader that fetches relevant memories on-demand. Use memory_context or retrieve_memories APIs to pull semantic, temporal, and entity-based memories into the current task context.

Yes, numpy is required as a dependency to implement these layered memory architectures. The memory system uses it to support consolidation processes and semantic retrieval for persistent agent state across sessions.