Latent Personal Memory: Represent personal memory as dynamic soft prompts
LPM addresses the need for efficient, scalable personalization of frozen LLMs, offering strong gains in accuracy and memory efficiency for user-specific tasks.
Latent Personal Memory (LPM) introduces a framework for personalizing LLMs by representing user-specific history as compact latent slots that generate dynamic soft prompts. On PersonaMem v1, LPM outperforms LoRA by 8.8% and Prompt Tuning by 54.4% in accuracy while reducing KV-cache usage by over 64x, and matches LoRA accuracy on LoCoMO with 120x fewer parameters.
Personalizing large language models (LLMs) requires encoding long-term, user-specific behavioral patterns in a way that is computationally efficient, scalable, and compatible with a frozen base model. We present Latent Personal Memory (LPM), a scalable framework that represents user-specific history as a compact, persistent matrix of N latent slots, that are interpretable. A shared cross-attention projection network maps these slots into dynamic, input-conditioned soft prompts that are prepended to the input of a frozen LLM. We evaluate LPM on PersonaMem v1 and LoCOMO benchmarks across Qwen3-1.7B, 4B, and 8B backbones. Results demonstrate that LPM outperforms LoRA and Prompt Tuning by up to 8.8% and 54.4% in overall accuracy respectively on PersonaMem v1, while reducing KV-cache usage by over 64x. On LoCoMo, LPM matches LoRA accuracy with 120x fewer trainable parameters. We also show that the efficiency of LPM grows with context length and outperforms full-context at 128K context length.