One-Line Summary

Proactive memory keeps agents on track in long tasks while new work clarifies attention trade-offs, brings text-to-music on device, and teaches video models to reason across frames.

Research Papers

Proactive memory agent prevents 'state decay' in long tasks

A separate memory agent runs alongside a long-horizon action agent and decides when to inject a short, memory-grounded reminder, helping the main agent recall facts, subgoals, and prior failures that matter later. On Terminal-Bench 2.0 and τ^2-Bench, this proactive approach increases pass@1 by +8.3 and +6.8 percentage points, respectively, without modifying the underlying action agent. ¹

Rather than passively exposing a memory bank, the system treats memory as an intervention: it maintains a structured memory from the recent trajectory and chooses whether to speak up or stay silent. In ablations, selective intervention outperforms passive bank exposure, always-on injection, advisor-only guidance, and general retrieval. ¹

This design addresses context-window limits that bury decision-relevant state as trajectories grow — what the authors term “behavioral state decay” — and it drops in as a plug-and-play module with existing agent harnesses and frontier action agents. ¹

To explore open-weight memory policies, the team trains Qwen3.5-27B on the SETA dataset using Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO), improving validation reward and showing partial transfer to Terminal-Bench. ¹

Linear attention trade-offs and cross-layer routing, mapped

This study puts four recurrent linear-attention designs — DeltaNet, Gated DeltaNet, Kimi Delta Attention, and Gated DeltaNet-2 — in a shared “recurrent memory” view and systematically compares training speed and quality. Experiments span 350M-parameter models trained for 15B tokens, plus larger DeltaNet runs at 1.3B and 3B parameters. ²

Within the 350M, 15B-token sweep, Kimi Delta Attention paired with the Muon optimizer reaches the lowest final validation loss; a pure Gated DeltaNet stack with AdamW attains the highest normalized training throughput; hybrid stacks generally reduce loss at a throughput cost; and Muon consistently lowers final validation loss relative to AdamW when architecture is matched. The authors also introduce Cross-Layer Value Routing (CLVR), forwarding write values between layers, which yields modest loss reductions for DeltaNet and Gated DeltaNet. ²

The reported speed numbers cover training throughput and iteration time; the authors do not provide an empirical inference-speed benchmark. The paper also compares optimizers and learning rates and measures how runtime scales with sequence length. ²

A compact, quantized runtime brings text-to-music on device

aria is a small, dependency-free native runtime that runs the complete Stable Audio 3 text-to-music pipeline on ordinary Graphics Processing Units (GPUs), on Central Processing Unit (CPU)-only machines, and even on a Raspberry Pi 5 — with no Python or deep-learning framework underneath. ³

Its core result is a careful quantization study: eight-bit precision shows no measurable quality loss across three independent measures (prompt adherence, overall audio quality, taste preservation) while sharply reducing memory and running fastest on GPU; four-bit adds a small, bounded quality cost but shrinks the footprint enough to fit the 1.2-billion-parameter model on an 8GB Pi 5. Compared with the official implementation, aria matches or exceeds generation speed and starts about seven times faster. ³

Because the runtime owns every internal tensor, it also exposes “activation steering” — a lightweight control to nudge what the model generates — which the authors showcase for “sonic seasoning.” The code is public, positioning aria as a practical base for on-device semantic audio in Internet-of-Sounds settings. ³

Teaching video generators to reason across frames

OpenCoF introduces a 17K-video reasoning dataset spanning 11 task families and fine-tunes a video model, Wan-CoF, to study whether diverse temporal supervision improves Chain-of-Frame (CoF) reasoning — a counterpart to Chain-of-Thought (CoT) where reasoning plays out over connected frames. Across four video-reasoning benchmarks, Wan-CoF records sizable gains over the Wan2.2-I2V-A14B baseline. ⁴

Beyond data, the authors add explicit “visual” and “textual” reasoning tokens to capture low-level cues and high-level semantic priors, then analyze where these tokens help across model depth, denoising steps, space, and time. They release the dataset, model, and code to support further work on reasoning-oriented video generation. ⁴

Open Source & Repos

LiteLLM: an OpenAI-format gateway to 100+ model APIs

LiteLLM is an open-source Software Development Kit (SDK) and proxy “AI gateway” that lets teams call more than 100 Large Language Model (LLM) Application Programming Interfaces (APIs) — including Bedrock, Azure, OpenAI, Vertex AI, Cohere, Anthropic, SageMaker, Hugging Face, vLLM, and Nvidia NIM — using a single OpenAI-compatible interface with built-in cost tracking, guardrails, load balancing, and logging. ⁵

In release v1.91.2 (Jul 11, 2026), the project highlights supply-chain security: all Docker images are signed with cosign, and releases are signed with the same key introduced in commit 0112e53. The repository also advertises one-click deploy options to Render, Railway, AWS CloudShell, and Google Cloud Shell for self-hosting. ⁵

Why It Matters

Long-horizon agents frequently forget decision-critical details as trajectories grow; proactive memory reframes “memory” as a selective intervention, recovering accuracy (+8.3 and +6.8 points pass@1 on two benchmarks) without changing the action agent itself. ¹

Alongside this, a clear map of linear-attention trade-offs, a quantized native audio runtime that fits a 1.2B model on an 8GB Pi, and new mechanisms for video reasoning signal maturing engineering patterns: choose efficient attention, push generation on device, and make intermediate reasoning state explicit. ²

What to Try This Week

1. LiteLLM quickstart: Read the README and try routing your app’s calls through the OpenAI-compatible gateway for cost tracking and logs. https://github.com/BerriAI/litellm
2. Proactive memory paper: Skim the abstract and ablations in “Remember When It Matters” to see how selective reminders beat always-on guidance. https://arxiv.org/abs/2607.08716