LMem: Lossless Token-Level Code Compression Exploiting AI Accent

A 47-entry dictionary designed by an AI agent (Claude Opus 4.6) compressed a 489-token Python file to 56 tokens — an 88.5% reduction, fully lossless. This paper makes three primary contributions. First, we report AI accent (formally: AI Lexical Convergence), the naming patterns that AI models converge upon through training to produce "readable code." AI accent is observed across three layers — variable names, class names, and file names — and is responsible for 82% of compression power. Analysis of an 18-file corpus (15,765 tokens) shows that the current word-level vocabulary covers 57.2% of tokens, and that code elements beyond the vocabulary can be compressed through granularity expansion of the dictionary, as demonstrated by the 88.5% file-optimized result. Second, we propose LMem, a dictionary compression method that exploits AI accent: a lossless token-level code compression that replaces recurring code patterns with single Unicode symbols processed as 1 BPE token, with runtime dictionary cost eliminated through LoRA internalization into model weights. Third, we analyze the safety asymmetry created by this technique — a model that has internalized LMem can produce output opaque to humans — and propose monitoring metrics. LoRA internalization on a 0.8B model (Qwen3.5) achieved 95% token accuracy (improved from initial 86%, targeting 100%).