COMA: Adaptive Context Budget Management for Memory-Augmented RAG via Query-Aware Zone Allocation, Memory-Conditioned Retrieval, and Cognitive Memory Consolidation

Most RAG pipelines do the same thing: embed the query, grab top-k chunks, stuff them into the context window until it's full, call the LLM. Nobody asks whether the retrieval zone should get 55% of the budget or 30%. Nobody checks whether the chunk you just retrieved is already sitting in working memory from two turns ago. Nobody models what the agent should remember across turns versus what it should actively retrieve. COMA fixes all three. It formalizes context assembly as a constrained budget optimization problem across five zones, extends Maximal Marginal Relevance to condition on agent memory state rather than just the current selection, and implements a three-tier memory system (sensory buffer, working memory, long-term with decay) inspired by the Atkinson-Shiffrin cognitive model and Ebbinghaus forgetting curves. Real results on GPT-4.1-mini: 37% fewer tokens, 2× recall improvement, 5.4× context efficiency gain over naive RAG. Full code and benchmarks released open source.