digest: 2026-05-18

_meta/claims-ledger.md

@@ -81,3 +81,11 @@
 - **2026-05-16** [arxiv:2605.15156] CLAIM: MeMo (MIT/NUS — Solar-Lezama, Rus, Chen, Low) encodes new knowledge into a dedicated memory model parallel to a frozen LLM; reports robustness to retrieval noise, no catastrophic forgetting, compatibility with open and proprietary LLMs without weight access, **inference cost independent of corpus size**; evaluated on BrowseComp-Plus, NarrativeQA, MuSiQue (abstract reports "strong" without specific numbers) | source: https://arxiv.org/abs/2605.15156 | status: open ; awaits concrete benchmark numbers and code/weight release; structurally complements [arxiv:2605.13941] EvolveMem (storage architecture vs. configuration evolution)
 - **2026-05-16** [arxiv:2605.13880] CLAIM: PREPING (KAIST — Sung Ju Hwang cluster; Proposer + Solver + Validator triple for pre-deployment synthetic-practice memory construction) — **2.99× lower deployment cost on AppWorld**, **2.23× lower on BFCL v3** vs. online memory construction; final task performance competitive with offline- or online-experience playbook baselines; covers MCP-Universe | source: https://arxiv.org/abs/2605.13880 | status: open ; "warm the cache before the user shows up" pattern for new MCP-server bootstrap; converges with [arxiv:2605.06614] SkillOS curated skill repo and [arxiv:2605.12493] LongMemEval-V2 AgentRunbook-C
 - **2026-05-16** [arxiv:2605.15077] CLAIM: AsyncFC (UC Berkeley — Joseph Gonzalez, Huanzhi Mao BFCL co-author) decouples LLM decoding from function execution; LLMs natively reason over symbolic futures (placeholders for unresolved tool results); no fine-tuning required; significantly reduced end-to-end task time across function-calling and SWE benchmarks while preserving accuracy (abstract conceptual, single headline number not surfaced) | source: https://arxiv.org/abs/2605.15077 | status: open ; runtime pattern composable on top of [arxiv:2605.10913] Shepherd substrate; directly applicable to existing MCP harnesses; open question on branching semantics when resolved future contradicts planned downstream branch
+- **2026-05-18** [arxiv:2605.16045] CLAIM: RecMem gates LLM-based memory extraction on observed *recurrence* of semantically similar interactions (subconscious embedding-only layer until clusters form); semantic refinement recovers fine-grained facts; reduces memory-construction token cost of three SOTA memory systems by **up to 87%** while *exceeding* their accuracy | source: https://arxiv.org/abs/2605.16045 | status: open ; attacks the cost axis of [arxiv:2605.12477] MEME (file-based on Opus 4.7 ~70× baseline cost) and [arxiv:2605.12493] LongMemEval-V2 AgentRunbook-C; open question whether recurrence-only consolidation evades the [arxiv:2605.12978] consolidation-decay failure mode or merely defers it
+- **2026-05-18** [arxiv:2605.16233] CLAIM: FORGE (population-based, no weight updates; reflection converts failed trajectories into reusable textual rules / few-shot examples, broadcasts best-performing memory across population between stages) reaches **1.7–7.7× over zero-shot**, **29–72% over isolated Reflexion**, **major-failure rate ~1%** on a 30-step CybORG CAGE-2 network-defense task; rules representation ~40% fewer tokens than alternatives; reproduced across 4 model families (Gemini, Grok, Llama, Qwen) | source: https://arxiv.org/abs/2605.16233 | status: open ; **10th memory-mechanism axis** alongside SkillOS / Belief Memory / Constant-Context Skill / MEMOREPAIR / DeMem / EvolveMem / MeMo / PREPING / DimMem — population-broadcast is mechanically distinct from all of them; open question whether broadcasting the best memory inherits [arxiv:2605.08060] Memory Curse in multi-agent settings
+- **2026-05-18** [arxiv:2605.15759] CLAIM: DimMem (atomic, typed, self-contained memory units with structured fields; dimension-aware retrieval + selective recall) — **81.43% LoCoMo-10**, **78.20% LongMemEval-S**, **24% per-query token reduction on LoCoMo**; a fine-tuned **Qwen3-4B** running DimMem exceeds LightMem + GPT-4.1-mini on both benchmarks; code released | source: https://arxiv.org/abs/2605.15759 | status: open ; concrete proof that a 4B local model is sufficient to drive structured-memory loops; converges with [arxiv:2605.15184] (grep beats vector) on "structure beats embedding sophistication"; open question whether typed schema survives [arxiv:2605.14498] GroupMemBench multi-party regime
+- **2026-05-18** [arxiv:2605.15315] CLAIM: LaMR (multi-rubric latent reasoning for coding-agent context pruning; decomposes relevance into *semantic evidence* + *dependency support* via specialized CRFs + MoE gating; multi-rubric labels derived from AST analysis, no human annotation) wins **12/16 multi-turn head-to-head comparisons** across SWE-Bench Verified / SWE-QA / LCC / LongCodeQA; **up to 31% token savings** on multi-turn tasks; **up to +3.5 Exact Match** single-turn | source: https://arxiv.org/abs/2605.15315 | status: open ; complementary to [arxiv:2605.15184] grep-vs-vector finding — grep wins as memory backend, LaMR-style structured pruning wins on file-selection; concrete production pattern for Claude Code-class harnesses
+- **2026-05-18** [arxiv:2605.16217] CLAIM: Argus (Searcher + Navigator roles, Navigator maintains shared evidence graph and verifies what's missing; 35B-A3B MoE backbone; RL-trained Navigator, standard ReAct Searcher; scales 1→64 parallel Searchers without retraining) — **+5.5 points avg with single Searcher across 8 benchmarks**, **+12.7 with 8 parallel**, **86.2 BrowseComp with 64 parallel** claimed to surpass every proprietary deep-research agent benchmarked; Navigator reasoning context **<21.5K tokens** even at 64 Searchers | source: https://arxiv.org/abs/2605.16217 | status: open ; **major leap over open-source SOTA** — vs. [arxiv:2605.05191] LongSeeker 61.5 BrowseComp and [arxiv:2605.04036] OpenSeeker-v2 46.0 BrowseComp; awaits independent reproduction at matched compute (64-Searcher run is compute-heavy); validates the planner-as-graph-manager pattern that [arxiv:2604.27221] Web2BigTable and [arxiv:2605.05007] Uno-Orchestra were pointing at
+- **2026-05-18** [arxiv:2605.15514] CLAIM: RoPE provably loses both (a) locality bias — failure probability approaches 0.5 (random) as context grows — and (b) consistency in token relevance ranking in long contexts; increasing RoPE base parameter trades token discriminability for positional discriminability and *cannot preserve both*; multi-head / multi-layer architectures shown insufficient to overcome | source: https://arxiv.org/abs/2605.15514 | status: open ; mechanistic foundation for [arxiv:2605.12922] "Channel-transition" / GAR, [arxiv:2605.10828] "First Drop of Ink", [arxiv:2605.12978] consolidation-decay — all RoPE-based long-context degradation patterns now have a provable architectural cause
+- **2026-05-18** [arxiv:2605.15777] CLAIM: SaaS-Bench (106 tasks × 23 deployable SaaS systems × 6 professional domains; multimodal long-horizon with verification checkpoints) — strongest evaluated computer-use agent completes **<4% of tasks end-to-end**; failure modes concentrated in planning, state tracking, cross-application context maintenance, and error recovery | source: https://arxiv.org/abs/2605.15777 | status: open ; widens the gap previously visible in [arxiv:2605.10787] ComplexMCP (<60%) and [arxiv:2605.12481] ToolCUA (46.85% OSWorld-MCP); honest baseline for enterprise-CUA deployment expectations
+- **2026-05-18** [arxiv:2605.14386] CLAIM: Darwin Family — training-free evolutionary weight-space merging via 14-dim adaptive genome (component + block level) + MRI-Trust Fusion (per-layer importance with learnable trust) + Architecture Mapper (Transformer × Mamba SSM crossover); **Darwin-28B-Opus** (Qwen3.6-27B hybrid linear/full attention, 64 layers 3:1 linear:full, Apache 2.0) reports **GPQA Diamond 74.75% single-shot greedy → 83.84% majority-vote@8 → 88.89% adaptive ensemble (~20× compute)**; ranks #6 of 1,252 models evaluated; Darwin-27B-Opus variant 86.9% | source: https://arxiv.org/abs/2605.14386 + https://huggingface.co/FINAL-Bench/Darwin-28B-Opus | status: open ; first credible "training-free merge keeps up with training" data point at this scale; comparable size class to Qwen3.6-35B-A3B → local-fit on M5 Max at Q4–Q6; headline 88.89% is Stage-3 ensemble not raw capability — 74.75% single-shot is the honest production comparable

_meta/processed.json

@@ -1,4 +1,12 @@
 [
+  { "id": "arxiv:2605.16045", "url": "https://arxiv.org/abs/2605.16045", "title": "RecMem: Recurrence-based Memory Consolidation for Efficient and Effective Long-Running LLM Agents", "date_covered": "2026-05-18" },
+  { "id": "arxiv:2605.16233", "url": "https://arxiv.org/abs/2605.16233", "title": "FORGE: Self-Evolving Agent Memory With No Weight Updates via Population Broadcast", "date_covered": "2026-05-18" },
+  { "id": "arxiv:2605.15759", "url": "https://arxiv.org/abs/2605.15759", "title": "DimMem: Dimensional Structuring for Efficient Long-Term Agent Memory", "date_covered": "2026-05-18" },
+  { "id": "arxiv:2605.15315", "url": "https://arxiv.org/abs/2605.15315", "title": "LaMR: Context Pruning for Coding Agents via Multi-Rubric Latent Reasoning", "date_covered": "2026-05-18" },
+  { "id": "arxiv:2605.16217", "url": "https://arxiv.org/abs/2605.16217", "title": "Argus: Evidence Assembly for Scalable Deep Research Agents", "date_covered": "2026-05-18" },
+  { "id": "arxiv:2605.15514", "url": "https://arxiv.org/abs/2605.15514", "title": "RoPE Distinguishes Neither Positions Nor Tokens in Long Contexts, Provably", "date_covered": "2026-05-18" },
+  { "id": "arxiv:2605.15777", "url": "https://arxiv.org/abs/2605.15777", "title": "SaaS-Bench: Can Computer-Use Agents Leverage Real-World SaaS to Solve Professional Workflows?", "date_covered": "2026-05-18" },
+  { "id": "arxiv:2605.14386", "url": "https://arxiv.org/abs/2605.14386", "title": "Darwin Family: MRI-Trust-Weighted Evolutionary Merging for Training-Free Scaling of Language-Model Reasoning", "date_covered": "2026-05-18" },
   { "id": "arxiv:2605.15184", "url": "https://arxiv.org/abs/2605.15184", "title": "Is Grep All You Need? How Agent Harnesses Reshape Agentic Search", "date_covered": "2026-05-16" },
   { "id": "arxiv:2605.14498", "url": "https://arxiv.org/abs/2605.14498", "title": "GroupMemBench: Benchmarking LLM Agent Memory in Multi-Party Conversations", "date_covered": "2026-05-16" },
   { "id": "arxiv:2605.13941", "url": "https://arxiv.org/abs/2605.13941", "title": "EvolveMem: Self-Evolving Memory Architecture via AutoResearch for LLM Agents", "date_covered": "2026-05-16" },