Graph-Augmented Recursive Language Models for Personal Knowledge Systems

% ============================================================ We present \textbf{Cog-RLM}, a graph-augmented recursive language model architecture for personal knowledge systems that achieves 90.3\% accuracy on a comprehensive 103-question evaluation spanning ten cognitive dimensions, using a stock 3-billion parameter model with zero fine-tuning and zero inference cost. Our system extends the Recursive Language Model (RLM) paradigm~\citep{zhang2025rlm} with three novel contributions: (1)~a local knowledge graph providing relationship-aware context retrieval via breadth-first traversal, (2)~a hybrid decomposition classifier that selectively triggers recursive processing only for multi-hop queries, reducing latency by 48\% versus always-on decomposition, and (3)~a persistent three-layer retrieval architecture combining static knowledge blocks, sentence-transformer embeddings, and graph traversal. We evaluate across ten dimensions---recall, reasoning, temporal, counterfactual, adversarial, generalization, consistency, precision, negation, and inference---demonstrating that architectural augmentation of small models outperforms fine-tuned models with $4\times$ more parameters by a factor of $5.4\times$ on domain-specific knowledge tasks. Our ablation study isolates the contribution of each component: RAG alone provides +66\% over fine-tuning, graph traversal adds +5\%, and RLM decomposition contributes +5\%, with the combination yielding multiplicative gains. The complete system runs on consumer hardware (Apple M4 Mac Mini, 16GB RAM, \$600) and processes queries in 1.0--12.5 seconds.