54 — Context-compression literature and market delta (2024–2026)

Companion to the headroom deep-dive (53): the internet re-sweep for context-compression projects and academic work the dossier's 03/46/51/52 do not already cover, run to answer "what else is out there, and did we miss anything?" Two parallel research efforts fed this file — a market sweep of shipping tools and a clean-room academic literature pass across eight threads — each independently cross-checked against primary sources. Research conducted 2026-06-15; every external claim carries a source in the ledger; tiers per the dossier scheme (T1 shipped+reproduced / T2 peer-reviewed / T3 community-replicated / T4 speculative). This is the compression-layer analogue of file 46's fresh-literature role.

TL;DR

• The single most useful synthesis is the cache-safety axis (§C): output brevity is cache-neutral; input compression is cache-breaking unless done at write-time on new content. Everything else sorts under this. On a hosted, already-caching Claude, a compressor that recompresses the prompt prefix must beat ~10× (pure prefix) / ~5.5× (mixed) just to break even against the 0.1× cache-read price — confirmed by a 358-run Claude Sonnet 4.5 RCT where aggressive compression raised cost 1.8% (arXiv 2603.23525) and by "Don't Break the Cache" (arXiv 2601.06007).
• The frontier moved to code-domain, hosted-viable, write-time compression that improves SWE-bench — refuting file 46 D's "no new compressor safe for code." Squeez (arXiv 2604.04979: 92% tool-output removal at 0.86 recall, write-time Unix pipe, cache-safe), AgentDiet (arXiv 2509.23586: Claude 4 Sonnet 64.5%→66.5%, input −40–60%, and the only paper that nets out its own +5–15% compressor cost), SWEzze/OCD (arXiv 2603.28119: +5–9.2% resolution at ~6×), SWE-Pruner (Claude Sonnet 4.5 70.6%→72.0%), LongCodeZip (ASE 2025: training-free, 5.6×, no loss).
• Stars are noise in this niche. A GitHub search found 63 repos created since Dec 2025 with >25k stars; the flagship "compression" repos show abnormal star:watcher ratios (RTK 62,461★/146 watchers = 428:1; CodeGraph 49,439/115; headroom 28k/95) on 3–5-month-old repos. Rank by evidence, forks, and open-issue activity, never stars.
• Most "compression" is selective inclusion, not payload compression. The flashiest multipliers (82×–528×, 120×, 99.1%) come from querying an index instead of reading whole files against a worst-case grep baseline — real, but the file-51/52 lever, not LLMLingua-style ratios; and several headline "98%" figures are bytes, not tokens.
• The peer-reviewed standouts are contrarian and code-relevant. the-complexity-trap (NeurIPS'25, JetBrains): cheap observation masking matches LLM summarization at half the cost on SWE-bench — and implicitly debunks the summarize-history compaction most agents ship. ACON (arXiv 2510.00615): 26–54% peak-token at >95% accuracy, gradient-free for hosted APIs. OpenHands batched condensation: the only widely-shipped cache-aware compaction.
• Independent headroom measurements land at ~47–50% (Miya-Gadget 47.5%, HN ~50%), below the 60–95% headline (full treatment in 53).
• Soft-prompt / embedding compression (gist/ICAE/500x/xRAG/PISCO/Cartridges) is confirmed a dead end on hosted Claude — every method needs embedding-input or KV-write access no hosted API exposes. Only Selective Context (a hard-prompt method) reaches a hosted API, at ~2×.

A. Market re-sweep — context-compression tools the dossier had not cataloged

Excludes everything already in 03/46/51/52 (caveman family, fff, claude-mem, aider map, LiteLLM, ccusage, TOON, LLMLingua, codedb, CodeGraff, Serena, the named context engines, vector DBs, KV-eviction families, CAG, Mem0). All numbers vendor self-reported unless tagged otherwise; none reproduced locally.

Five skeptical filters (apply to every row)

1. Stars are a PR artifact (see TL;DR) — use forks/issues/watchers and evidence instead.
2. Selective inclusion ≠ payload compression. "Query an index" multipliers are the file-51 lever against a worst-case baseline; do not compare them to fixed-prompt compression ratios.
3. Evidence is almost all vendor self-report. Peer-reviewed exceptions are rare and mostly NL-domain.
4. NL→code transfer is unproven for most compressors (benchmarked on NQ/TriviaQA/HotpotQA, not SWE-bench).
5. Cache breakage is the silent cost-killer. History-rewriting compaction/proxies can destroy the 10× cache-read discount they nominally beat (Anthropic docs: "Tools that drop or summarize older messages… invalidate cache hits"). Only a handful engineer around it.

Condensed market table (load-bearing rows)

(The fuller per-category catalog — proxies, libraries, tool-output hooks, code-index "selective inclusion" tools, semantic caches, compaction harnesses, memory systems, and RAG-chunk compressors — and every star/fork figure with its watcher-ratio flag is preserved in the source ledger below; the rows above are the ones that move a decision.)

Credible challengers ranked by evidence (not stars)

1. the-complexity-trap (JetBrains-Research, NeurIPS'25, MIT) — the most credible artifact in the sweep: peer-reviewed, code-specific (SWE-bench Verified), and contrarian — cheap observation masking matches expensive LLM summarization at ~half the cost. A method, not a drop-in, but it is the only rigorous code evidence here and it debunks the summarize-history compaction that Cline/Goose/Hermes ship.
2. OpenHands Condenser (MIT) — best-evidenced shipping compaction with a real SWE-bench number and the only harness cache-aware by design (batched condensation).
3. ACON (Microsoft, peer-reviewed) — the most conservative, therefore most believable, agent-context number: 26–54% peak-token at >95% accuracy; gradient-free so it runs against hosted APIs.
4. llmtrim (AGPL) — the most caching-honest proxy: explicitly compresses "without touching the cached prefix" and discloses its measurement limits. A more transparent, narrower headroom competitor.
5. claw-compactor (MIT) — strongest library: zero-LLM, deterministic, reversible (RewindStore = an open, inspectable CCR), AST-aware that never renames identifiers, honest 36.3% with an apples-to-apples ROUGE-L baseline vs LLMLingua-2.
6. flightlesstux/prompt-caching (MIT) — orthogonal and the one tool whose savings are provably real because it only uses Anthropic's documented cache_control; maximizes caching instead of risking it.

Cautionary tales (high hype, weak substance): RTK (62k stars, but its own issue tracker says the hook raises cost 18% and bypasses Claude Code's permission prompts — output-rewriting hooks are cache-unsafe); the "82×–528× / 120× / 99.1%" code-index tools (selective inclusion vs worst-case grep, not compression — the file-51 lever); and any soft-token RAG method (xRAG/REFRAG/PISCO/OSCAR), which cannot run against the Anthropic API at all (§B thread 1).

B. Fresh literature (2024–2026), by thread

The meta-finding: the field moved from NL-QA soft-prompt compression (a dead end for hosted APIs) to code-domain, agent-trajectory, hosted-compatible compression evaluated on SWE-bench — and several of these improve accuracy while cutting tokens.

• Thread 1 — Soft-prompt / embedding compression: dead for hosted Claude. Gist (26×), ICAE (4×), 500xCompressor (6–480×, 62–73% retention), xRAG (1 token/doc), AutoCompressors, COCOM, PISCO (16× @ 0–3% loss), Cartridges (38.6×), LLoCO (30×) — all compress into embeddings/KV the model must be trained to read; surveys say plainly they are "impractical for API-based proprietary LM services" (arXiv 2410.12388). Correction to the dossier's K1: 500xCompressor writes per-layer KV — the "works without fine-tuning" framing is misleading for hosted use. Only Selective Context (a hard-prompt method, ~2×) reaches a hosted API.
• Thread 2 — AST/code-aware compression: the new, hosted-viable, accuracy-positive frontier. The Perplexity Paradox (arXiv 2602.15843): LLMLingua on code causes "Function Identity Collapse" — 86.1% of failures are NameError — fixed by deterministic signature injection (+34pp; NameError 86.1%→6.1%), which is exactly headroom's CodeAwareCompressor design. SWEzze/OCD (arXiv 2603.28119): AST-aware, ~6×, SWE-bench resolution +5.0–9.2%. LongCodeZip (ASE 2025): training-free, model-agnostic, 5.6× without loss on code. CodeMEM (arXiv 2601.02868): AST memory, no fine-tuning, +12.2% instruction accuracy. REPOFUSE: signature-only "rationale context." All run as preprocessing on hosted Claude but are query-conditional → cache-breaking per instance (so: compress at ingestion, not in the prefix).
• Thread 3 — Tool-output / observation compression: the richest vein, several measured on Claude. AgentDiet (arXiv 2509.23586): the canonical "separate cheap summarizer" pattern and the only one that accounts for its own cost (+5.2–14.8%) — net Claude 4 Sonnet 64.5%→66.5%, input −39.9–59.7%. ACON (arXiv 2510.00615): gradient-free, hosted-API-compatible, peak-token −26–54.5%. M2 (arXiv 2603.00503): training-free, Claude-3.7 +12.5%/−30.3% tok, Claude-4 +5.5%/−39.1%. CaT/SWE-Compressor (arXiv 2512.22087): SWE-bench 49.8%→57.6%. First-party: Claude API automatic compaction (beta, default 150k-token trigger; invalidates the cached prefix unless you breakpoint the system-prompt end + the compaction block; iteration tokens billed but excluded from top-level usage) and Claude Code's microcompact (no-LLM redundant-tool-output trimming every turn — the cheap, always-cache-safe layer).
• Thread 4 — Reversible / retrievable compression: paging composes with caching; context editing breaks it. MemGPT/Letta (DMR 93.4% vs 35.3% summary baseline — but a full-context baseline hit 94.4%, so paging buys little when everything fits), Anthropic memory tool + just-in-time context (+39%/84% on a 100-turn web eval, append-style, cache-friendly). The exception: Anthropic context editing explicitly invalidates the cached prefix on each clear (clear_at_least_input_tokens exists precisely to make the re-cache worth it). Headroom CCR and claw-compactor's RewindStore are the buildable reversible+cache-aware tools.
• Thread 5 — Learned-importance selection. Squeez (arXiv 2604.04979): the standout — task-conditioned tool-output pruning, 92% removal at 0.86 recall, runs as a write-time Unix pipe (pytest | squeez 'find the failure'), cache-safe, code-domain (BM25 only 0.22 recall on tool output — retrieval fails on logs). SWE-Pruner (GitHub Ayanami1314/swe-pruner; arXiv ID tentative): 0.6B skimmer, Claude Sonnet 4.5 70.6%→72.0%, tokens −23–38%, but mid-stream → cache-breaking. Provence (ICLR'25): strong but NL-QA only, never code-evaluated.
• Thread 6 — Context rot: the accuracy argument for compression. Chroma's "Context Rot" (2025-07, 18 models incl. Claude Opus 4 / Sonnet 4 / 3.7 / 3.5 / Haiku 3.5): reliability falls as input grows even on trivial retrieval, with Claude Opus 4 / Sonnet 4 showing the most pronounced gap (they abstain under ambiguity); all 18 did better on a shuffled than a logically-ordered haystack. Plus "Lost in the Middle" (U-shaped, >30% drop) and RULER (17/17 long-context models degrade). ACE (ICLR'26): evolving playbooks with incremental updates avoid "context collapse," +10.6% on agents, token −83.6% — but base model is DeepSeek, not Claude. Cognition's "Don't Build Multi-Agents" is the counterpoint for coding: prefer a single linear agent + a dedicated compression LLM.
• Thread 7 — THE CRUX (input compression vs caching). No hosted-API scheme makes per-turn input compression cache-compatible. Break-even ≈ 10× pure-prefix / 5.5× mixed. Hard evidence: "Don't Break the Cache" (arXiv 2601.06007: caching saves 41–80%, put dynamic content last) and the 358-run Claude 4.5 RCT (arXiv 2603.23525: aggressive compression raised cost 1.8%). Every cache-aware compressor (CacheBlend EuroSys'25 Best Paper; KVzip NeurIPS'25) is self-host-only. The only hosted-Claude-safe input moves are (a) offline compress-once → cache (the CAG pattern, file 46 FL1) and (b) write-time observation compression before content enters the cached prefix.
• Thread 8 — Output brevity vs input compression. Output is billed ~5× input and is cache-neutral, so brevity is strictly the safer lever. Chain of Draft (arXiv 2502.18600, tested on Claude 3.5 Sonnet): sports 92.4% output cut +4.1pp, GSM8K ~79% cut −4.4pp, latency −48%. TALE (ACL'25): 68.9% at <5% loss. Token Complexity (arXiv 2503.01141): the guardrail — every task has a minimal-token floor below which accuracy collapses. Brevity-hierarchy (arXiv 2604.00025): verified to exist (file 03 flagged it unverified) but a trap — solo/unreviewed, +26.3pp is a cherry-picked 7.7% subset, no Claude, no code; do not propagate "+26 points."

C. The unifying synthesis — a cache-safety classification of every compression move

This is the table the dossier was missing: sort compression by where it acts, and cache-safety falls out deterministically. It is the practical decision rule for a hosted-Claude operator.

D. Corrections and refinements to the dossier

• New caveat — stars are not adoption in the compression niche. The dossier's market files (03/51/52) cite star counts as an adoption signal; in this sub-field that is now invalid (PR-spike inflation, abnormal watcher ratios). Rank by evidence/forks/issues. Applied retroactively to file 53's headroom/caveman star figures.
• Refute file 46 D ("no new lossy compressor both user-reachable on hosted Claude and safe for code"). Superseded by Squeez, AgentDiet, SWEzze, SWE-Pruner, LongCodeZip — code-domain, hosted-viable, accuracy-neutral-or-positive — provided they act at write-time/ingestion, not on the cached prefix. The Perplexity Paradox supplies the mechanism (signature injection beats perplexity pruning on code).
• Sharpen the break-even (record 19 / file 46 FL3). State it as ~10× on a fully-cacheable prefix, ~5.5× on a mixed prompt, with the cacheable-fraction assumption explicit — the single most attackable number in the dossier's chain, now backed by a Claude-specific RCT.
• Correct the caveman brevity citation (file 03 record 01 / K1). arXiv 2604.00025 is real but NL-only, unreviewed, cherry-picked; the defensible "brevity helps accuracy" evidence is Chain-of-Draft on Claude 3.5 Sonnet (+4.1pp at −92% output), a modest effect consistent with the dossier's existing read.
• Correct the soft-prompt framing (K1). 500xCompressor/ICAE/xRAG are not hosted-usable "without fine-tuning" — they need embedding-input or per-layer-KV channels; the family is self-host-only.
• Promote output brevity above input compression in the recommendation hierarchy on cache-neutrality grounds (§C).

E. jackin' implications

The compression layer sorts cleanly onto the cache-safety classification (§C), and the jackin' recommendation follows from it:

1. Bake the cache-safe set, in order: output brevity (already shipped via the caveman family) → write-time observation compression (pilot headroom MCP mode or a hook equivalent like the-complexity-trap masking / chop / Squeez-style filtering) → native cache-maximization (/cd, stable prefixes, cache_control discipline) → offline compress-once for stable corpora (the CAG pattern).
2. Do not put a whole-prompt compressor in a jackin' container's hot path — it is a cache-bust risk, a CompressionAttack surface (file 46 FL3), and on already-caching Claude Code it usually loses (the RTK issue tracker is the worked cautionary tale: a hook that raised cost 18%).
3. Prefer peer-reviewed, code-domain, cache-aware tools when piloting: the-complexity-trap (observation masking), OpenHands-style batched condensation, ACON. Rank candidates by evidence, never by stars.
4. Reuse the file-51 validation harness with cache-read continuity as a first-class metric (already added in 53's harness): accept a compressor only if it beats the existing hooks+code-intel arm on tokens-per-solved-task at equal quality and preserves cache_read.
5. Memory layer: pick one reversible, cross-agent store if needed (headroom memory, Letta, or Zep/Graphiti — the only one with paired before/after token counts), and demand net-accounting before trusting it.

Verification ledger

All accessed 2026-06-15. Independent measurements and peer-reviewed work flagged inline; everything else is vendor self-report.

Independent headroom measurements: Miya-Gadget benchmark (47.5%, RAG 0%) — miyagadget.page/en/blog/2026/06/03/headroom-ai-context-compression-benchmark-en (via reader proxy; site 403s direct) · HN "~50%" — news.ycombinator.com/item?id=46663757 · negative evidence (HN story index has zero hits for the tool) — hn.algolia.com/api/v1/search?query=headroom&tags=story · press churn origin — theregister.com/ai-ml/2026/05/31/netflix-wiz-creates-app-to-slash-ai-bills-then-open-sources-it

Market tools: github.com/fkiene/llmtrim · github.com/juyterman1000/entroly · github.com/borhen68/TokenTamer · github.com/open-compress/claw-compactor · github.com/jia-gao/leanctx · github.com/rtk-ai/rtk (Issue #582, #886) · github.com/mksglu/context-mode · github.com/AgusRdz/chop · github.com/atlassian-labs/mcp-compressor · github.com/flightlesstux/prompt-caching · github.com/OpenHands/OpenHands · github.com/microsoft/acon · github.com/JetBrains-Research/the-complexity-trap · github.com/getzep/graphiti · github.com/zilliztech/GPTCache · github.com/Compresr-ai/Context-Gateway · github.com/colbymchenry/codegraph · github.com/tirth8205/code-review-graph · github.com/DeusData/codebase-memory-mcp · github.com/letta-ai/letta · github.com/firecrawl/firecrawl-mcp-server

Literature: Squeez arXiv 2604.04979 · AgentDiet arXiv 2509.23586 · ACON arXiv 2510.00615 · the-complexity-trap arXiv 2508.21433 (NeurIPS'25 DL4Code) · Perplexity Paradox arXiv 2602.15843 · SWEzze/OCD arXiv 2603.28119 · LongCodeZip ASE 2025 (arXiv 2510.00446) · M2 arXiv 2603.00503 · CaT arXiv 2512.22087 · CodeMEM arXiv 2601.02868 · "Don't Break the Cache" arXiv 2601.06007 · Claude 4.5 compression RCT arXiv 2603.23525 · Chain-of-Draft arXiv 2502.18600 · TALE arXiv 2412.18547 (ACL'25 Findings) · Token Complexity arXiv 2503.01141 · brevity-hierarchy arXiv 2604.00025 · prompt-compression survey arXiv 2410.12388 · 500xCompressor arXiv 2408.03094 · xRAG arXiv 2405.13792 · PISCO arXiv 2501.16075 · Cartridges arXiv 2506.06266 · Chroma "Context Rot" trychroma.com/research/context-rot · ACE arXiv 2510.04618 · CacheBlend (EuroSys'25) arXiv 2405.16444 · KVzip arXiv 2505.23416 · Anthropic compaction platform.claude.com/docs/en/build-with-claude/compaction · Anthropic context engineering anthropic.com/engineering/effective-context-engineering-for-ai-agents

• Companion deep-dive: 53-headroom-and-context-compression.md. Prior market/lit files: 03, 46, 51, 52.

Caveats on this file's own data

• Star counts are point-in-time (2026-06-15) and unreliable as quality signals in this niche (§A filter 1); reported only because adoption was asked about, with the watcher-ratio flag.
• Two arXiv IDs could not be independently confirmed and may be mis-transcribed or pre-registration placeholders: SWE-Pruner's (cited as 2601.16746) and a structural-graph paper (2603.27277, also referenced in file 51's ledger). The GitHub repos behind them are real and star-verified; treat those two specific identifiers as tentative pending a direct arXiv check.
• This file synthesizes two parallel research sweeps plus direct primary-source verification of every load-bearing claim; per the dossier's standing rule, no headline percentage here is a banked saving until reproduced on jackin' tasks via the file-51/53 validation harness at equal quality.