1\section{Related work}2\label{sec:related}34\paragraph{Text and prompt compression.}5Shortening the text a receiver reads is the oldest form of inter-agent compression. Beyond truncation and extractive selection,6learned compressors drop tokens with a classifier trained on distilled data \citep{pan2024llmlingua2}; LLMLingua-2 is7task-agnostic by construction, which places it in our \emph{question-blind} regime. Abstractive summaries and structured8extraction are the dominant representations in deployed agent frameworks; we treat them as first-class baselines rather than9as foils, tuning their prompts on development data only.1011\paragraph{KV-cache quantization, eviction and compaction.}12KIVI \citep{liu2024kivi} quantizes keys per channel and values per token to 2 bits with a full-precision residual window, motivated13by outlier channels in keys. Eviction methods select positions to keep: H$_2$O \citep{zhang2023h2o} accumulates attention mass14during prefill; SnapKV \citep{li2024snapkv} scores prefix positions by the attention of an observation window at the end of the15prompt, pooled and selected per head; PyramidKV \citep{cai2024pyramidkv} allocates decreasing budgets to higher layers;16StreamingLLM \citep{xiao2024streamingllm} keeps attention sinks and a recent window and \emph{re-assigns} positions within the17cache. Attention Matching \citep{zweiger2026compaction} compacts a cache into fewer positions by solving for keys and values that18reproduce attention outputs on reference queries. Our audit (Appendix~\ref{app:audit}) records for each method whether its19selection observes the question: SnapKV- and PyramidKV-style selection are question-aware by construction when the prompt ends20with the question, which matters for the comparison with reusable, question-blind messages.2122\paragraph{Cache reuse and transfer.}23Prefix caching reuses an identical token prefix within one checkpoint; it is exposed by local runtimes \citep{mlxlm2026,hf2026kvcache}24and, as a billing abstraction, by hosted APIs \citep{anthropic2026caching}. Transfer across checkpoints requires a mapping:25Cache-to-Cache \citep{fu2026c2c} trains a fuser that injects a sharer's cache into a receiver that still prefills the source;26LatentMAS \citep{zou2026latentmas} exchanges full caches and latent thoughts among agents that share the same architecture;27SwaT \citep{alabdulkareem2026swat} trains a cross-family translator and combines it with compaction, but releases no checkpoint.28Concurrent work on cross-model cache bridging (see the audited list in \texttt{literature/sources.csv}) is active; we make no29novelty claim about shared latent spaces and run only a same-family linear mapper as a bounded, fully disclosed experiment.3031\paragraph{Evaluation.}32LongBench \citep{bai2024longbench} supplies public long-document QA with F1 scoring; RULER \citep{hsieh2024ruler} generates33controlled retrieval tasks. We combine both ideas: public subsets (QASPER, 2WikiMultihopQA) for external validity, and a new34procedural generator of private facts (identifiers, amounts, dates, negations, timestamped revisions, unanswerable questions) so that35communication dependence can be tested without pre-training contamination.36