memory-architecture

Vector-only memory fails because it treats memory as a single component rather than a layered system. A coherent agent memory stack requires distinct semantic, episodic, procedural, and working memory layers with defined cross-layer interfaces to avoid architectural gaps and missing entity anchors.

Designing a layered agent memory architecture involves defining implicit, semantic, episodic, procedural, and working memory responsibilities. You must establish clear cross-layer flows for entity anchoring, procedure discovery, and working-memory routing to connect retrieval and reasoning paths without architectural gaps.

Route queries by applying a cognitive taxonomy to map retrieval requests to the correct memory layer. This working-memory routing directs traffic to appropriate semantic, episodic, or procedural paths, ensuring the agentic reasoning system retrieves context from the correct memory architecture layer.

Common memory architecture anti-patterns include using vector-only storage, performing episodic dumps without entity anchors, causing working memory overflow, and lacking decay policies. These failure modes disrupt cross-layer integration and prevent proper procedure discovery from episodic traces to procedural DAGs.

Different agent types require different layer compositions. The memory architecture provides design templates for scenarios like personal assistants and procedural autonomous agents, allowing you to select the correct combination of semantic, episodic, procedural, and working memory layers for your specific use case.

Procedural memory connects to episodic memory through the procedure discovery flow. This cross-layer integration converts episodic traces into procedural DAGs, allowing the memory-augmented agent to anchor entities and discover repeatable procedures within its memory architecture.