digest: 2026-05-18

[JustinPerea]

Summary

8 items today, covering 7 of the 10 priority areas.

Memory (4 items — the dominant cluster this run):

• RecMem (2605.16045) — recurrence-triggered consolidation; -87% memory-construction token cost vs. three SOTA systems while exceeding their accuracy. Directly attacks the cost axis MEME exposed at ~70× on Opus 4.7.
• FORGE (2605.16233) — population-broadcast textual memory, no weight updates; 1.7–7.7× over zero-shot on CybORG CAGE-2 across 4 model families. 10th distinct memory-mechanism axis in the ledger.
• DimMem (2605.15759) — atomic typed memory units; 81.43% LoCoMo / 78.20% LongMemEval-S; a fine-tuned Qwen3-4B beats LightMem + GPT-4.1-mini. Code released. Strong local-fit signal.

Context engineering (2 items):

• LaMR (2605.15315) — multi-rubric (semantic-evidence + dependency-support) context pruning for coding agents; wins 12/16 multi-turn head-to-heads on SWE-Bench Verified / SWE-QA / LCC / LongCodeQA; -31% tokens / +3.5 EM.
• RoPE provably fails (2605.15514) — locality bias and token discriminability cannot both be preserved in long contexts; mechanistic foundation for the channel-transition / first-drop-of-ink / consolidation-decay findings already in the ledger.

Orchestration / reasoning:

• Argus (2605.16217) — Searcher + Navigator + shared evidence graph; 86.2 BrowseComp with 64 Searchers on a 35B-A3B MoE, Navigator context stays <21.5K tokens. Major jump over prior open-source SOTA (LongSeeker 61.5, OpenSeeker-v2 46.0) — flagged as a conflict pending reproduction at matched compute.

Evaluation:

• SaaS-Bench (2605.15777) — 106 tasks × 23 real SaaS systems; top computer-use agent <4% end-to-end. Widens the gap visible in ComplexMCP and ToolCUA.

Local / frontier models:

• Darwin-28B-Opus (2605.14386) — training-free evolutionary merge on Qwen3.6-27B (Apache 2.0, hybrid linear/full attention). 74.75% GPQA Diamond single-shot → 88.89% via Stage-3 adaptive ensemble (~20× compute). Quants already on HF; fits M5 Max 64GB comfortably at Q4–Q6.

Conflicts flagged

1. Argus 86.2 BrowseComp vs. prior open-source SOTA — large leap (LongSeeker 61.5, OpenSeeker-v2 46.0, Tongyi DeepResearch 43.2). Mechanism (Navigator-managed evidence graph + parallel Searcher dispatch) is novel rather than scale-only, but the 64-Searcher headline is compute-heavy and prior baselines were typically run at lower parallelism. Awaits independent reproduction at matched compute budgets.

Sources

arXiv cs.AI / cs.CL / cs.MA recent (May 17–18), HF Daily Papers, HF blog, Anthropic news + research, DeepMind blog. OpenAI research returned 403 (retry next run). Lab-blog round-up: no new technical drops from Anthropic research since 2026-05-14; AlphaEvolve impact post on DeepMind was already-tracked theme.

Test plan

• Confirm inbox/2026-05-18.md renders cleanly in Obsidian
• Spot-check the 8 arXiv links resolve
• Verify _meta/processed.json parses (89 total entries, 8 new for today)
• Verify _meta/claims-ledger.md appended 8 new claim entries

https://claude.ai/code/session_0146cLi2ZfHHMDtpgJtmJiwy

---

Generated by Claude Code