Ingested from a three-cycle multi-stream survey (Streams A–E, evidence tiers S–D, coverage 2024-01 → 2026-07) on LLM session caching, KV cache compression, context compaction, session lifecycle management, and autonomous decision policies. This file is the nexus's extraction: the findings we act on, the layers we explicitly do not act on, and the survey's gap register mapped to ECM mechanisms. Companion spec:
docs/specs/ecm/spec.md.
The survey covers a seven-layer dependency stack. Eidolons is host-mediated (Claude Code, Codex CLI, Copilot, Cursor, OpenCode): the bottom four layers belong to providers and serving stacks and are out of scope for nexus tooling. They are retained here because they explain why the top-layer contracts must look the way they do.
| Layer | Examples | Nexus posture |
|---|---|---|
| Architectural | MLA (~75% KV reduction, DeepSeek V2+) | Observe only |
| Serving system | PagedAttention (SOSP'23), RadixAttention (SGLang) | Observe only; deployment note for junction sandbox loops |
| Cross-engine tier | LMCache (arXiv:2510.09665, 15× throughput), NVIDIA Dynamo | Observe only |
| KV compression | KVzip (NeurIPS'25 Oral), MiniKV (86% @ >98.5% acc), ChunkKV | Observe only — but its failure modes inform the pin-set contract |
| Provider API caching | Anthropic cache_control (0.10× reads), OpenAI auto, Gemini implicit, DeepSeek disk |
Act: prefix-stability contract (ECM §cache-discipline) |
| Session management | Server-side compaction (Anthropic compact-2026-01-12, OpenAI /responses/compact), context editing, memory tool |
Act: compaction triggers, externalize-before-compact, handoff (ECM core) |
| Agent policy | Utilization thresholds, cost heuristics — no formal model exists (Gap G1) | Act: mechanical decision table (ECM §policy) |
- Wang et al. (arXiv:2601.15300, 2026): open-model F1 collapses 45.5% at 40–50% of max context (Qwen2.5-7B, 0.55 → 0.30); five-method cross-validated.
- EMNLP Findings 2025: input length alone degrades performance 13.9–85% across 5 models even with perfect retrieval and whitespace-only distractors.
- Chroma (18 frontier models, B/C-tier): every model degrades with context growth; past ~50% of window, degradation correlates with distance-from-end (recency bias).
- Claude Code's own auto-compact default is 95% of window, with 60–75% the
recommended aggressive override range (
CLAUDE_AUTOCOMPACT_PCT_OVERRIDE).
Taken: ECM's zone ladder — GREEN < 50% / AMBER 50–75% / RED 75–90% / CRITICAL ≥ 90%. Compact at RED, never wait for the host's 95% emergency floor. The 50%-vs-77% source tension is resolved by scope: the 40–50% collapse is open-model evidence; frontier hosts degrade more gracefully — so 50% opens the advisory zone, 75% is the mechanical trigger.
- Chen et al., ACL 2026 (arXiv:2510.00231): under multi-instruction prompting, KV eviction bias causes certain instructions to be completely ignored, and demonstrably leaks system-prompt content. Negligible-loss claims from single-instruction benchmarks do not transfer.
- Compaction is the application-layer analog of eviction: server-side compaction is lossy by design; exact tool arguments, file contents, identifiers are not guaranteed to survive.
Taken: ECM's pin set — a declared set of context elements (cortex digest, refusal tables, ESL enforcement mode, tier map, active criteria SHAs) that must survive every lossy operation verbatim, with post-op verification and repair by re-injection. This is the same failure class ACL 2026 documents, addressed at the layer we control.
- Anthropic + OpenAI docs and practitioner evidence converge: anything that must survive compaction exactly (paths, symbols, decision IDs, plan state) goes to persistent storage before the trigger — never trust the summary for exact recall.
- Anthropic's memory tool + context editing integration formalizes the pattern: warn-before-clear, write-to-memory, then clear.
Taken: ECM's externalize-before-compact contract, implemented on
CRYSTALIUM's existing surface (plan_checkpoint for execution state, commit
for episodic notes, ingest for ECL envelopes). Mandatory when crystalium is
installed; warn-and-degrade when absent.
- All four provider caches key on exact prefix match (Anthropic breakpoints, OpenAI/Gemini automatic, DeepSeek disk). Cache reads cost 0.10–0.50× base input; a stable 100k-token prefix is the single highest-ROI optimization (up to 90% input-cost and 85% latency reduction).
- LMCache production finding: context truncation halves prefix-cache hit ratio — the standard cost-control move (truncate from the front) directly defeats the standard efficiency move (prefix caching).
Taken: ECM's cache-discipline contract: stable-prefix ordering (cortex → agent files → tool defs first; volatile injections appended last), never truncate-from-front, byte-stable generated sections (the existing idempotency + marker invariants already deliver this — ECM names why they are load-bearing), short per-prompt injections to bound cache-miss cost.
- Anthropic beta
compact-2026-01-12: configurable trigger (min 50k, default 150k), human-readable summary block,pause_after_compactionfor inject-resume, custom instructions replace the default summary prompt. Known limitation: with tools defined, the model may tool-call during summarization — instructions must prohibit it. - OpenAI
/responses/compact+context_management.compact_threshold: opaque encrypted compaction item (not inspectable — a real operational difference). - Claude Code: PreCompact hook fires before auto/manual compaction; auto-compact threshold overridable.
Taken: ECM triggers compaction through the host's surface (hooks, /compact,
API compaction where the harness is API-driven) and never re-implements it. The
PreCompact-class hook is where externalize-before-compact and pin re-injection
mechanize.
- Mem0 (platform-side automatic retrieval/injection) succeeds where Letta/MemGPT-class "agent remembers to call memory tools" fails — the same measured failure as our documentary cortex (DOSSIER-HARNESS-2026-06 §2).
Taken: the meter, the policy verdict, and the handoff recall are all hook-injected artifacts computed by the deterministic kernel — never capabilities the model must remember to request. This is why ECM's implementation is kernel verbs + hook recipes first, MCP tools a distant optional second.
Continue / prune tool results / compact / handoff-fresh / subagent-isolate /
graceful wrap-up. The survey's five-operation taxonomy plus budget ceiling
(pause_after_compaction + token counter) covers every observed production
pattern.
Taken: ECM's operation set is exactly this closed set — no open-ended "context strategies."
No S/A-tier study of 10+ compaction events exists; summary-of-summary loss is hypothesized (survey §5.4 H2) but unmeasured.
Taken: ECM takes the conservative position: compaction depth cap = 2. Beyond two compactions in one session, the only permitted lossy operation is handoff-fresh (externalize → brief → new session), which resets summary depth to zero with crystalium as the fidelity floor.
- KV compression families (eviction / quantization / merge / sharing / head-pruning; MiniKV, KVzip, ChunkKV, ClusterAttn, DMS, AsymKV, FlowMM): provider/serving-layer concerns. Only their pitfall literature (F2) crosses into ECM.
- MLA and serving-stack economics (PagedAttention, RadixAttention, LMCache, CacheBlend, Dynamo): relevant only as a deployment note — self-hosted junction sandbox loops with repeated large prompts benefit from SGLang+LMCache-class serving; not spec surface.
- Multi-tenant KV-sharing leakage (NDSS 2025) and DroidSpeak cross-model KV reuse: serving-layer; noted in ECM's security considerations only as the adjacent failure class.
| Survey gap | Survey status | ECM response |
|---|---|---|
| G1 — formal cache/compact/reset cost model | No S/A-tier work | Mechanical decision table over observable signals (ESL right-sizing doctrine: never LLM-discretionary). Not an optimization model — a deterministic policy with recorded provenance, auditable and revisable. |
| G2 — cross-session persistence quality | No S/A-tier work | Handoff brief (structured, ECL-enveloped, crystalium-ingested) + round-trip canary (successor session must recall the brief; measurable, CI-checkable). |
| G3 — compaction-boundary information leakage | No S/A-tier work | Trust-tier preservation: handoff briefs and compaction-survivor notes carry contains_tool_origin provenance when T3 content was in session scope; Dream's corroboration gate guards promotion; pins never include T3 content. |
| G4 — repeated-compaction fidelity | No S/A-tier work | Depth cap = 2, then handoff-fresh (F8). |
| G5 — leaked-material stream | Empty by absence | n/a |
- Kwon et al., PagedAttention, SOSP 2023, arXiv:2309.06180 — [S]
- Chen et al., Pitfalls of KV Cache Compression, ACL 2026, arXiv:2510.00231 — [S]
- Wang et al., Intelligence Degradation in Long-Context LLMs, arXiv:2601.15300 — [A]
- Du et al., Context Length Alone Hurts LLM Performance, EMNLP Findings 2025 — [S]
- Kim et al., KVzip, NeurIPS 2025 Oral, arXiv:2505.23416 — [S]
- Liu et al., LMCache, arXiv:2510.09665 — [A]
- Yao et al., CacheBlend, EuroSys 2025 Best Paper, arXiv:2405.16444 — [S]
- Anthropic Platform Docs — compaction (
compact-2026-01-12), context editing (context-management-2025-06-27), memory tool (memory_20250818), prompt caching — [B, first-party] - OpenAI API Docs —
/responses/compact,context_management.compact_threshold— [B, first-party] - DOSSIER-HARNESS-2026-06 (internal) — hook-surface matrix, inject-by-default verdict, Mem0-vs-Letta memory-injection evidence — [internal, verified]
Full survey (all tiers, all six subdomains, methodology, banned-source register) retained in the campaign records; this file is the actionable extraction.