Small adapters can provide frozen decoder-only LLMs with persistent latent-space memory that survives across separate sessions.
March 25, 2026
Original Paper
Trained Persistent Memory for Frozen Decoder-Only LLMs
arXiv · 2603.22329
The Takeaway
Traditionally, decoder-only models are stateless. This work shows that memory can be 'injected' via self-attention adapters, enabling models to retain knowledge or context across sessions without full fine-tuning or massive context windows.
From the abstract
Decoder-only language models are stateless: hidden representations are discarded after every forward pass and nothing persists across sessions. Jeong (2026a) showed that trained memory adapters give a frozen encoder-decoder backbone persistent latent-space memory, building on the lateral-memory framework of Jeong (2026b,c). Here we ask whether the same principle transfers to the decoder-only setting, where no cross-attention pathway exists and memory must enter through self-attention alone. We a