memory-systems

Memory systems provide the necessary persistence layer for AI agents to maintain continuity, track entity states, and perform complex reasoning across multiple sessions. This skill guides developers through the architecture of layered memory, moving from simple volatile context windows to sophisticated temporal knowledge graphs. It is designed for engineers and architects building stateful production agents who need to balance retrieval accuracy with system complexity. By evaluating production-grade frameworks including Mem0, Zep/Graphiti, Letta, LangMem, and Cognee, this skill helps select the optimal storage backend based on requirements like entity consistency, multi-hop reasoning, and latency constraints. It emphasizes the importance of reliable retrieval over tool complexity, drawing on benchmark data such as LoCoMo and LongMemEval to guide architectural decisions.

• Framework Evaluation: Compare architectural trade-offs between vector-store based memory (Mem0), temporal knowledge graphs (Zep/Graphiti), self-editing tiered storage (Letta), and semantic graph pipelines (Cognee).

• Layered Design: Implement strategies for working memory, short-term session state, entity-specific registries, and long-term knowledge retention using graph databases or document stores.

• Retrieval Optimization: Configure multi-hop reasoning, relationship traversal, and temporal filtering to enable agents to perform time-travel queries and maintain context across long-running interactions.

• Benchmark-Driven Selection: Utilize data from DMR, LoCoMo, and HotPotQA to assess memory performance, latency, and reasoning correctness before committing to a production stack.

• Use this skill when you need to solve the 'cold start' problem in multi-session agents or when standard RAG pipelines fail due to lack of entity relationships.

• Start with the shallowest memory layer (e.g., file-system or cache) before escalating to full graph-based persistent storage.

• Integrate with agent workflows to maintain user preferences and domain-specific knowledge across disparate conversation threads.

• Consider input constraints such as ingestion-time processing costs in Cognee versus the ease of deployment in Mem0.

• Focus on maintaining high-signal tokens within the context window, using memory tools only for information that exceeds the immediate attention budget.