\section{Related work}
\label{sec:related}

\paragraph{Text and prompt compression.}
Shortening the text a receiver reads is the oldest form of inter-agent compression. Beyond truncation and extractive selection,
learned compressors drop tokens with a classifier trained on distilled data \citep{pan2024llmlingua2}; LLMLingua-2 is
task-agnostic by construction, which places it in our \emph{question-blind} regime. Abstractive summaries and structured
extraction are the dominant representations in deployed agent frameworks; we treat them as first-class baselines rather than
as foils, tuning their prompts on development data only.

\paragraph{KV-cache quantization, eviction and compaction.}
KIVI \citep{liu2024kivi} quantizes keys per channel and values per token to 2 bits with a full-precision residual window, motivated
by outlier channels in keys. Eviction methods select positions to keep: H$_2$O \citep{zhang2023h2o} accumulates attention mass
during prefill; SnapKV \citep{li2024snapkv} scores prefix positions by the attention of an observation window at the end of the
prompt, pooled and selected per head; PyramidKV \citep{cai2024pyramidkv} allocates decreasing budgets to higher layers;
StreamingLLM \citep{xiao2024streamingllm} keeps attention sinks and a recent window and \emph{re-assigns} positions within the
cache. Attention Matching \citep{zweiger2026compaction} compacts a cache into fewer positions by solving for keys and values that
reproduce attention outputs on reference queries. Our audit (Appendix~\ref{app:audit}) records for each method whether its
selection observes the question: SnapKV- and PyramidKV-style selection are question-aware by construction when the prompt ends
with the question, which matters for the comparison with reusable, question-blind messages.

\paragraph{Cache reuse and transfer.}
Prefix caching reuses an identical token prefix within one checkpoint; it is exposed by local runtimes \citep{mlxlm2026,hf2026kvcache}
and, as a billing abstraction, by hosted APIs \citep{anthropic2026caching}. Transfer across checkpoints requires a mapping:
Cache-to-Cache \citep{fu2026c2c} trains a fuser that injects a sharer's cache into a receiver that still prefills the source;
LatentMAS \citep{zou2026latentmas} exchanges full caches and latent thoughts among agents that share the same architecture;
SwaT \citep{alabdulkareem2026swat} trains a cross-family translator and combines it with compaction, but releases no checkpoint.
Concurrent work on cross-model cache bridging (see the audited list in \texttt{literature/sources.csv}) is active; we make no
novelty claim about shared latent spaces and run only a same-family linear mapper as a bounded, fully disclosed experiment.

\paragraph{Evaluation.}
LongBench \citep{bai2024longbench} supplies public long-document QA with F1 scoring; RULER \citep{hsieh2024ruler} generates
controlled retrieval tasks. We combine both ideas: public subsets (QASPER, 2WikiMultihopQA) for external validity, and a new
procedural generator of private facts (identifiers, amounts, dates, negations, timestamped revisions, unanswerable questions) so that
communication dependence can be tested without pre-training contamination.