Trained Persistent Memory for Frozen Decoder-Only LLMs

TL;DR

This paper explores methods to enable persistent memory in frozen decoder-only language models like GPT-2, demonstrating that architectural priors significantly influence memory retention at low capacity, with all methods converging at higher capacity.

Contribution

It adapts six memory methods to frozen GPT-2, revealing the importance of architecture in persistent memory and establishing it as a general paradigm across transformer models.

Findings

Cross-attention, Hebbian, and slot write methods achieve 7-18% memory scores at low capacity.

All six methods converge at higher capacity, indicating architectural influence.

Memory methods improve knowledge retention in frozen decoder-only models.

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 adapt six methods -- prefix, parallel cross-attention, KV extension, Hebbian memory, context-gated branch, and slot-based sparse write -- to a frozen GPT-2, training only a small adapter $\theta_{mem}$. The write rule is shared; only the read injection changes from decoder cross-attention to self-attention KV prefix or parallel branch. On LoCoMo we find a striking inductive-bias dichotomy: at $1\times$ capacity, three methods with strong architectural priors -- cross-attention (M.2), Hebbian (M.4), and slot write (M.6) -- achieve retained-memory scores of $7-18\%$ and knowledge gains $\Delta K$ of $7-10$, while the other three fail ($< 0.4\%$). At $10\times$ capacity all six converge, showing the gap is architectural, not fundamental. Together with the encoder-decoder results of Jeong (2026a) and the brain-inspired modules of Jeong (2026b,c), these findings establish persistent latent-space memory as a general paradigm spanning major transformer families.