Token-Efficient Conversational History Management using Sparse Semantic Patch Memory

Large Language Model (LLM) deployments face increasing constraints related to context window limitations, inference cost, and redundant conversational history storage. Conventional raw-token history retention results in inefficient memory utilization and unstable reasoning under bounded token budgets. This work introduces Sparse Semantic Patch Memory (SSPM), a semantics-first memory architecture that decomposes conversational history into compact, utility-scored semantic patches rather than monolithic token sequences. A DeepSeek-inspired extraction module identifies entities, constraints, decisions, structured artifacts, and code fragments, which are normalized into patch representations and retained through budget-aware Top-K optimization. We formulate patch selection as a constrained optimization problem under token budgets and demonstrate through proof-of-concept experiments that SSPM reduces stored tokens by ~37% while preserving critical constraint and decision continuity. This study establishes a systems-level foundation for token-efficient conversational memory management and opens directions for scalable, budget-aware LLM reasoning architectures.