LLM agents operating continuously face a fundamental constraint: the finite context window limits memory between sessions. We present .ava, a compressed symbolic notation achieving 2.65x compression ratio in tokens compared to French natural language. Unlike vector memories, .ava is human-readable, editable, and version-controllable. We formalize its grammar, report fine-tuning experiments on Qwen2.5-1.5B (6 GB VRAM, 78.5% token accuracy), and propose a cognitive architecture including the concept of curvature. Also available in French: .ava : Une notation symbolique compressée pour la mémoire persistante d'agents IA
Cros et al. (Sun,) studied this question.