UQO Working Papers · WP30 · v1.0
Beyond Text: Cache Compression and Knowledge Transfer in Multi-Agent Systems
A Reproducible Study of Fidelity, Communication Cost, and Repeated Handoffs
- Year
- 2026
- First release
- October 2026
- Latest revision
- October 2026
- Publication status
- Working paper
Lossless handoffs are interchangeable; every lossy choice depends on model and regime
Across 13,280 trials on two 3B models, text, zstd text and native or 8-bit caches give identical answers; 4-bit quantization is free for Llama and fatal for Qwen, and attention-based eviction keeps 78–98% of private-fact accuracy with the question but only 12–19% without it.
Abstract
Agents built on language models hand information to one another as text, summaries, structured evidence, shared-memory references or, when they share a checkpoint, attention state (the KV cache). This paper asks under which task, model, cache, network and reuse conditions each representation preserves task-relevant information at the lowest end-to-end cost. The study is executed by an autonomous research operator (Claude Code) on Apple-Silicon hardware with the MLX runtime, whose caches are inspectable tensors, so that all byte counts are measured on serialized payloads and all latencies are synchronized wall-clock spans.
Twenty-two conditions (truncation, extractive selection, summaries, structured evidence, metered retrieval, a lossless codec, LLMLingua-2, native and serialized caches, packed 8/4-bit quantization and four eviction rules) are measured on Qwen2.5-3B and Llama-3.2-3B across private-fact ledgers, LongBench QA and executable tasks, in question-blind and question-aware regimes, over 13,280 trials under a frozen protocol with cluster bootstrap and Holm-corrected tests. Further experiments measure two-node cache transfer over a 1 GbE LAN, relays of 1–8 handoffs with an evaluator-only fact ledger, fan-in with conflicting evidence, a same-family learned KV mapper, hosted receivers and a development-fitted router.
Lossless representations are interchangeable in quality and differ only in cost, whose ranking inverts between cold and reused regimes; on the measured link, moving an 8-bit cache beats re-reading the text at every length. Lossy choices are model- and regime-dependent: 4-bit quantization is free for Llama and fatal for Qwen; attention-based eviction keeps 78–98% of private-fact accuracy with the question and 12–19% without it; no eviction survives a code specification. At a 512-token budget prose summaries beat structured evidence (38–44 vs 12–15%); relays are nearly lossless after the first hop and invent almost nothing. A ridge KV mapper transfers no content while producing coherent abstention, and the router halves amortized latency at equal quality when questions are unknown. Cross-family cache transfer was not executed.
Key findings
13,280
trials across 22 conditions on two 3B models
78–98%
private-fact accuracy kept by question-aware eviction
12–19%
the same eviction without the question
38–44 vs 12–15%
prose summaries vs structured evidence at 512 tokens
1,676 vs 3,667 ms
8-bit cache transfer vs remote prefill at 8k tokens
−934 ms
router latency saving at equal quality, questions unknown
Scientific contributions
- A common harness for inter-agent handoffs
- Twenty-two text, cache and evidence representations are measured under one protocol with validated cache adapters, deterministic scorers, evaluator-isolated answer keys and document-level cluster bootstrap, over 13,280 trials.
- Lossless representations are interchangeable
- Text, its zstd encoding, the native cache, the reloaded cache and the 8-bit packed cache give identical answers; the choice is a cost question whose ranking flips between cold (text wins) and reused (cache wins) regimes.
- Lossy compression depends on model and regime
- 4-bit per-token quantization is free for Llama-3.2-3B and fatal for Qwen2.5-3B; attention-based eviction retains 78–98% of private-fact accuracy with the question and 12–19% without it; no eviction survives a code specification.
- Measured two-node break-even on 1 GbE
- On an M3 Ultra receiver, an 8-bit cache transferred over the LAN answers an 8k-token context in 1,676 ms against 3,667 ms for remote text prefill, so moving the cache already wins at about 2,000 tokens on this link.
- Relays lose at the first hop only
- Over 1–8 handoffs with an evaluator-only fact ledger, agents compress by omission rather than invention (about 0.05 unsupported additions per message); at a 512-token budget prose summaries carry more than structured evidence (38–44 vs 12–15%).
- A learned KV mapper that transfers nothing
- A same-family ridge mapper explaining 68% of the receiver's cache variance carries no content and yields coherent abstention (−75 points vs the receiver's own reading), an argument for compatibility manifests and negative controls in any cache-transfer protocol.
Methodology
Methodology (5)
Taxonomy of communication representations
Text compression, KV compression, cache reuse, cache transfer and parametric transfer are separated, and each method is tagged with its question-access regime, compatibility requirements and composability across sources.
Validated cache adapter layer
Identity round trip, prefix equivalence, position preservation, packed-byte accounting and manifest-checked loading are tested on a real 1.5B model before any performance number is recorded.
Frozen protocol and confirmatory statistics
Ten hypotheses were registered before any held-out outcome; six confirmatory comparisons are Holm-corrected, with document-level cluster bootstrap intervals; everything else is exploratory.
Two-node transfer and break-even model
Mac Studio M3 Ultra sender and receiver on a 940 Mbit/s LAN (1.2 ms RTT) compare cache transfer against remote text prefill and local recompute, with a measured per-token model of receiver prefill versus link time.
Relays, fan-in, mapper and router
Chains of 1–8 handoffs with metered repair, fan-in with conflicting evidence, a ridge KV mapper (Qwen2.5-1.5B to 3B) with an identity-transfer negative control, and a router fitted on development data and evaluated on held-out data against the best static policy.
Research data
- Procedural private-fact ledgers
- Synthetic documents with invented names, identifiers, amounts, dates, negations, timestamped revisions and unanswerable questions, generated to test communication dependence without pre-training contamination; answer keys are held by the evaluator only.
- LongBench QA subsets and executable tasks
- Public long-document QA (QASPER, 2WikiMultihopQA) scored with LongBench F1 for external validity, plus executable specification tasks simplified on development data before any test item was run.
- Per-attempt trial records
- 13,280 core trials plus relay, fan-in, two-node, mapper, hosted-API and router runs on Qwen2.5-3B and Llama-3.2-3B (bf16 mlx-community checkpoints), each with measured serialized bytes, synchronized latency spans and peak MLX memory, released as JSONL.
Reproducibility
- Harness src/wp30 with CLI (doctor, data prepare/validate, models probe, run, reproduce), pinned lock file and 20 correctness, leakage and scoring tests
- Protocol frozen with a hash before the test split existed; registered hypotheses and analysis plan in protocol/
- Every per-attempt record in runs/raw; make reproduce-analysis rebuilds all tables and figures with a provenance index and claim ledger
- Full LaTeX source of the 49-page paper; minimal reproduction profile re-runs a fixed subset and compares greedy outputs
Research artifacts
Version history
| Version | Date | Description | File |
|---|---|---|---|
| v1.0current | October 2026 | Current manuscript as published on www.spboucher.ai (revised edition after the 2026-10-17 scientific review for papers 1–17). | PDF · 49 p. |