memory-systems

Memory systems for agents use layered architectures combining working, short-term, and long-term memory layers plus temporal knowledge graphs. This Skill covers designing these layered models, integrating vector stores for retrieval, and implementing consolidation logic so agents maintain context and reasoning across multiple conversations and sessions.

Working memory holds immediate context for current tasks, short-term memory stores recent interactions, and long-term memory persists entity relationships and historical patterns. Temporal knowledge graphs add time-aware semantics so the system tracks when facts remain valid, enabling agents to distinguish outdated from current information.

Vector stores index embeddings for semantic retrieval, while graph memory structures entity relationships and temporal constraints. This Skill provides integration patterns to wire both into your agent's context pipeline, seed initial memories via provided scripts, and query across both storage types for cross-session reasoning.

Yes. Memory consolidation extracts user preferences and past interactions from individual conversations and stores them in long-term entity memory. Agents retrieve these consolidated patterns to personalize responses and maintain consistency, improving accuracy and relevance without reprocessing every prior exchange.

This Skill requires numpy for numerical operations on embeddings and memory vectors. Optional graph tooling extends temporal query capability, but the core layered memory and vector-store integration work with numpy as the foundation dependency.

Use temporal knowledge graphs when agents must reason about fact validity over time—for example, tracking user preferences that change, remembering event timestamps, or distinguishing stale from current information. Flat storage loses temporal context; graphs enable time-aware queries that improve accuracy in dynamic domains.