Text‑based memory and context compression enable reuse of past interactions. Resizing a continuous memory changes the input to a frozen LLM, coupling capacity allocation with readout. We introduce Hard‑Origin Adaptively Softened Memory (HasMem), where frozen hard‑prompt embeddings give a verifiable initial state. A controller adjusts memory widths, a Writer re‑encodes resized entries, and Reader together with Global provide readout adaptation and cross‑turn state sharing. In a reconstruction probe derived from the Multi‑Session Chat development split (535 questions), the main configuration attains lexical F1 of $95.3$ (a $4.4$‑point gain) while preserving $93.6\%$ of the hard reference’s framed memory positions. With comparable per‑question body budgets, six configurations with mean per‑entry retention $0.83\sim0.91$ surpass rule‑based re‑encoding by $8.0\sim23.6$ EM points. Holding model parameters fixed and using a rule target width ratio of $0.75$, Global’s EM improvement passes a Bonferroni‑corrected paired test over eight comparisons. On the 500 LongMemEval‑S questions, local lexical F1 rises from $3.4$ to $8.9$, and answer negative log‑likelihood drops from $12.257$ to $5.274$. The F1 gains accompany slight EM declines on both evaluations.
Review