Indexed Key Retrieval from LoRA Adapters for Continual Learning

This paper presents a mechanism for encoding personal knowledge into LoRA adapter weights with per-fact addressable recall. Each fact is assigned a unique string key, and the adapter learns to return the corresponding QA pair when prompted — enabling enumeration, verification, and lifecycle management of stored knowledge. A SimHash fingerprint registry provides hallucination detection at 8 bytes per key. The paper validates the core mechanism to 100 keys across three model families (Qwen 2.5 3B, Gemma 2 9B, Mistral 7B) on a single GPU (8 GB VRAM, QLoRA 4-bit), and reports both successes and revealing failures: incremental training without replay causes catastrophic forgetting (0/40 old keys after 5 cycles), and neither additive composition nor weight merging preserves indexed recall across adapters. Extended experiments including scaling to 550+ keys with 100% recall maintained, natural language generalization tracking, and a working home deployment are documented in the project repository: https://github.com/tiberius1701d/paramem