New Architecture Routes Agent Memory Queries for Enhanced Performance

A new architectural design called Supra Cognitive Modes (SCM) has been proposed to optimize how AI agents manage and access their memory. Agent memory workloads are diverse, ranging from simple factual lookups to complex multi-hop reasoning and broad synthesis over extensive historical data. SCM addresses this by dynamically mapping explicit or automatically selected query modes to appropriate retrieval and synthesis mechanisms, all operating on a unified data ingest substrate. The architecture features a frozen semantic classifier and runtime gates that direct queries to fused lexical and dense lookup, graph-based or iterative multi-hop processing, or stratified long-form synthesis. The underlying substrate integrates multi-granularity embeddings, extracted triples, and fact-version metadata. Evaluated on benchmarks like Long-term Conversational Memory (LoCoMo), MemoryAgentBench (MAB), and LongMemEval, SCM achieved strong results, indicating its effectiveness in handling diverse memory tasks and improving agent performance.