digest: 2026-06-15

[JustinPerea]

Summary

• 7 verified items, 3 conflicts/open-questions flagged.
• Priority areas covered: agent memory (digest: 2026-05-04 #1), context engineering (digest: 2026-05-06 #2), agent orchestration (digest: 2026-05-07 #3), tools/computer-use (digest: 2026-05-08 #5), evaluation (digest: 2026-05-12 #9), local-fit (digest: 2026-05-10 #7), coding agents (digest: 2026-05-11 #8).

Headline signals

• GitOfThoughts ([arxiv:2606.14470]) directly refutes the "memory helps" assumption. Across 5 substrates × 2 benchmarks × 2 model scales, memory does not reliably help on novel problems — only ≥0.8-similarity (near-duplicate) retrieval works. Tensioned against the entire memory-architecture cluster (MRAgent, AutoMEM, DMF, MAGE, EMBER, HORMA, TokenMizer) — flagged as primary conflict.
• StreamMemBench ([arxiv:2606.14571]) is the streaming-eval analogue: agents store and locally apply info but fail to translate to future behavior.
• AGMem ([arxiv:2606.14106]) — Action-Grounded Visual Memory delivers +33.3% on OSWorld vs full-image memory. Drop-in design pattern for computer-use agents.
• Parallel-Synthesis ([arxiv:2606.14672]) — KV-cache aggregation from parallel worker branches, matches/beats text on 7/9 datasets at 2.5–11× lower TTFT. Second KV-cache-as-communication paper in two days (composes with yesterday's Dense Latent Communication).
• HarnessX ([arxiv:2606.14249]) — third harness-layer paper in five days; typed primitives + substitution algebra + AEGIS evolution; +14.5% avg / +44% max across 5 benchmarks. Complements RAH (structure) + HarnessBridge (controller) + OrchRM (reward).
• WorkBench Revisited ([arxiv:2606.13715]) — clean 2-year longitudinal frontier data: GPT-4 March 2024: 43%/26% harmful → Claude Opus 4.8 June 2026: 89%/2.5% harmful.
• CacheRL ([arxiv:2606.14179]) — Qwen3-4B-Thinking trained to 92% process accuracy vs GPT-5 94% at ~100× less compute; RL improved stability but added little over strong SFT. Local-fit for M5 Max.

Conflicts flagged

1. GitOfThoughts vs the memory-architecture cluster — direct empirical tension; open question logged about realized similarity distribution on LoCoMo/LongMemEval/MemoryAgentBench.
2. Parallel-Synthesis vs Dense Latent Communication — complementary (parallel vs sequential), converging on cache-not-text thesis; open question about cross-family generalization.
3. HarnessX vs RAH vs HarnessBridge vs OrchRM — four complementary harness-layer primitives in five days, no head-to-head yet.

Not a slow-news day — arXiv June 12 backlog released overnight after Sunday's quiet, plus thin June 13–15.

---

Generated by Claude Code