Indian Patent Application & Intellectual Property Disclosure Invention: Surprise-Gated Persistent On-Device Memory with Importance-Weighted Decay (Aeon Core Architecture). Patent Office: Government of India — The Patents Act, 1970 & The Patents Rules, 2003 (Form 2 Provisional Specification). Inventor & Applicant: Er. Infant Nirmal Priyan J, Adrient Pvt Ltd, Bangalore, Karnataka, India. 1. Abstract & The Personal AI Thesis Conventional transformer-based large language models are amnesic: every interactive session reinitializes from identical static weights. Expanding context windows scales cost linearly or quadratically with history length, whereas persistent user memory at cloud scale poses severe privacy risks and prohibitive infrastructure costs. Aeon Core is an on-device, encrypted cognitive memory framework that continuously acquires personalized facts, patterns, and behavioral preferences. Memory formation is governed by surprise, retention is regulated by importance-weighted exponential decay, and consolidated episodic traces synthesize into high-order semantic representations during device idle cycles. 2. Core Mathematical Formulations Formula 1 — Surprise-Gated Memory Writes write(event) ⇔ surprise(event) > write_gate importance_i = α·surprise_i + (1 - α)·recency_i [α ≈ 0.6] Events enter episodic memory only when they deliver unpredicted information. Trivial or redundant inputs are bypassed automatically. Formula 2 — Importance-Weighted Exponential Decay retention_i(t) = exp( -λ·t / importance_i ) tombstone iff retention_i(t) < r_forget [r_forget = 0.05] Biological forgetting as a first-class feature: high-importance, frequently recalled traces persist indefinitely; transient noise evaporates. Formula 3 — The Certainty Signal (Honesty Calibration) confidence = f( self_stats[domain], retrieval_support, rollout_agreement ) if confidence < τ → explicit refusal or tool verification Certainty is derived from mathematical confidence accumulators rather than prompt vibes, eliminating confident hallucinations. Formula 4 — Offline Sleep Replay & Consolidation pattern_strength = count(episodes) × cosine_sim(z_episodic) promote_to_semantic iff pattern_strength ≥ τ_cluster During system idle, recurring episodic events are consolidated into compact semantic facts while raw sensory traces are pruned. 3. Empirical Proof: Zero Catastrophic Forgetting In dual sequential grammar transition tests (Environment A followed by Environment B, verified in native Rust testbeds): 100.00% Environment A Retention (8/8) 87.50% Environment B Acquisition (7/8) +87.50 pp Advantage Margin Over Control Baseline (12.50%)
