SynapseMem: Deterministic Temporal Belief Invalidation and Ebbinghaus Memory Consolidation for Edge Language Models

Deploying autonomous agentic workflows on resource-constrained edge hardware requirespersistent, adaptive memory. While Retrieval-Augmented Generation (RAG) based on vectorembeddings remains the standard paradigm for external memory, it suffers catastrophicallyfrom chronological state amnesia. When environmental configurations, system states, or userfacts mutate over time, semantic similarity search retrieves both obsolete and updated textpassages with equivalent proximity scores, inducing severe hallucinations in Small LanguageModels (SLMs ≤ 3B parameters). Furthermore, unstructured context appending introduceslinear prompt growth (O(N)), degrading local CPU inference throughput.In this work, we present SynapseMem, a zero-dependency, edge-first temporal beliefgraph memory engine. SynapseMem formalizes factual interactions into atomic belief tuples(s, p, o) managed via a deterministic Semantic Singularity Gate that automatically supersedesstale state variables without secondary LLM arbitration. We integrate an Ebbinghaus-inspiredexponential decay and synaptic reinforcement function to naturally prune conversationalentropy while cementing core facts. Evaluated against qwen2.5:3b on sequential stateinversion benchmarks, SynapseMem achieves 100% temporal state resolution, yields a 48.6%reduction in prompt context footprint, enforces a strictly bounded O(1) context overheadper entity query, and delivers a 58.3% reduction in query response latency compared to naivecontext accumulation.