Latent Personal Memory: Represent personal memory as dynamic soft prompts

The article introduces Latent Personal Memory (LPM), a framework for personalizing large language models (LLMs) by encoding user-specific behavioral patterns as a compact matrix of latent slots. LPM utilizes a cross-attention projection network to generate dynamic soft prompts that are prepended to the input of a frozen LLM, demonstrating superior performance on the PersonaMem v1 and LoCoMo benchmarks with Qwen3 models, achieving up to 8.8% and 54.4% accuracy improvements over LoRA and Prompt Tuning, while significantly reducing KV-cache usage. This approach enhances efficiency with increasing context lengths, making it a valuable method for practitioners looking to optimize personalization in LLMs.