Skip to content

Primary-source research ledger

Access date: 2026-11-21 UTC (2026-11-20 local). All lookups were read-only/free. No model provider, token-counting API, active agent session, or credential store was invoked.

Closeout refresh on the same UTC date re-opened the official MCP 2025-11-25 overview, Anthropic caching page, BLAKE3 specification repository and Unicode Braille chart. The chart again redirected to Unicode 18.0. This refresh confirms the cited source locations and scope; it does not establish provider behavior, mathematical proofs or redistribution permission.

Standards and mechanism checks

Source and version Checked section / supporting content (paraphrase) What it does and does not establish
MCP 2025-11-25 Overview: stateful JSON-RPC connections and capability negotiation Normative legacy contract; handler support alone is not lifecycle conformance
Anthropic prompt caching Automatic/explicit cache control, prefix identity, 5-minute and 1-hour TTLs Supports cache-policy review; no application hit rate measured
BLAKE3 specifications Maintainer specification/analysis repository Full hash/XOF/MAC design; not proof that a 48/64-bit prefix is collision free
Unicode Braille chart Redirects to Unicode 18.0, block U+2800–U+28FF; first page terms distinguish chart reproduction Establishes codepoints, not BPE tokens or general chart redistribution rights
xRAG v2 Abstract: trainable modality bridge from document embeddings into a frozen model Its results do not validate textual AHTL sketches without such a bridge
DuoAttention v1 Abstract: full KV for retrieval heads, smaller cache for streaming heads Model/cache-internal work, not evidence of preserved semantics for arbitrary symbolic syntax
Infini-attention v2 Abstract: compressive memory integrated into transformer attention Inspiration for local bundling must be separated from implementing the neural architecture
LLMLingua-2 v2 Abstract: distilled token classification with encoder, measured compression on named tasks Does not establish AHTL token ratios or reasoning quality
Cache Engineering v1 Abstract: serving-cache operators and engine evaluation Reports paper-specific 27.7% average cost reduction and 1.5x P95 speedup; not ADR's 4–8x TTFT claim
ACM v1 Title/abstract: Agentic Context Management, offload/query tools and post-training Confirms correction from “Cache” to “Context”; no demonstrated equivalence to icx wire caching
LCM v1 Title/date/abstract: summary DAG retaining original-message pointers Primary page says submitted 14 Feb 2026 despite identifier prefix. No fabricated date correction inferred
Fractal KV Archives v2 Abstract: symbols after quantization, lossless archive vs lossy quantizer Lossless symbolic archive does not imply lossless recovery from icx's 16-dimensional trajectory projection

Source summaries are deliberately short. Workstream reports carry more specific protocol, security, redistribution and mathematical checks near each finding.

Bibliographic identity pass: all 22 indexed arXiv records

Every abstract URL below resolved. Version is the latest version exposed by the page during this pass; this is not a claim that each frozen local PDF has that version. Local reference-PDF hashes, page counts and title metadata are in reference-pdf-metadata.json.

Record Latest observed version Identity/result scope
2610.28577 v1 Cache Engineering for Agents; abstract checked
2607.23809 v1 Agentic Context Management; abstract checked
2607.07144 v2 Fractal KV-Cache Archives; abstract checked
2605.04050 v1 Lossless Context Management; abstract checked
2512.24601 v3 Recursive Language Models; identity/version checked
2510.11967 v1 Context-Folding; identity/version checked
2510.04618 v3 Agentic Context Engineering; identity/version checked
2502.12110 v11 A-MEM; identity/version checked
2501.13956 v1 Zep; identity/version checked
2501.00663 v1 Titans; submission date 31 Dec 2024
2410.15704 v1 Residual vector quantization; single author/title checked
2410.10819 v1 DuoAttention; abstract checked
2406.10774 v2 Quest; identity/version checked
2406.02069 v4 PyramidKV; identity/version checked
2405.14831 v3 HippoRAG; identity/version checked
2405.13792 v2 xRAG; abstract checked
2404.14469 v2 SnapKV; identity/version checked
2404.07143 v2 Infini-attention; abstract checked
2403.12968 v2 LLMLingua-2; abstract checked
2401.03462 v3 Activation Beacon; identity/version checked
2310.08560 v2 MemGPT; identity/version checked
2307.06945 not captured ICAE; title/resolution checked; version not independently transcribed

Explicit research limits

  • The 22-record pass checks identity and selected abstract claims; it does not independently reproduce all paper experiments or re-review every line of all papers.
  • Maintained website pages can change; versioned MCP/arXiv links are used where possible. The Unicode unversioned URL redirected to 18.0 during this audit. Source identity is not license permission.
  • Historical ev_* binary/SQLite evidence identifiers were read as provenance claims, not treated as current schema or authorization. Private databases were not opened to validate them.
  • Non-arXiv historical bibliography and every external prior-art repository were not exhaustively revalidated. A local filesystem path to ai-memory is not independently accessible public prior art.
  • No Anthropic tokenizer, real cache counters, live model reasoning effectiveness, repository branch-protection configuration or patent filing records were verified. These remain follow-up procedures, not positive assurance.
  • Where current primary sources conflict with a frozen paper or historical record, preserve both with version/date and explain the change; do not silently rewrite immutable reference material.