Paper: In-Process Retrieval as Extended Working Memory for Language Agents
Paper: In-Process Retrieval as Extended Working Memory for Language Agents
A new arXiv paper proposes ‘Memory in the Loop’, which moves knowledge retrieval inside the agent’s reasoning loop. Instead of querying external memory once per turn, the agent reads and writes memory on every step. Experiments show this architecture significantly improves performance on long-horizon tasks requiring continuous context reference.
Lilian Weng Summarizes 35 Papers on Harness Engineering for RSI
OpenAI’s Lilian Weng published a summary of 35 papers on ‘Harness Engineering for Recursive Self-Improvement (RSI)’. The review covers prompt engineering, reward design, self-correction mechanisms, and other approaches to enable AI systems to improve themselves recursively. It is a rare deep-read material for the week.
ChatGPT Voice Mode Upgraded to GPT-5.5, Can Delegate Hard Tasks
OpenAI upgraded the model powering ChatGPT’s voice mode to a new version (referred to as GPT-5.5). Preview access is available on the iPhone app. The new model is significantly more fluent and accurate in voice interactions. Critically, it can automatically delegate harder tasks—requiring web search, deep reasoning, or heavy computation—to a more powerful backend, while keeping voice interactions low-latency.
Vercel Agent Debuts: An AI Agent That Can Touch Production
Vercel announced an expansion of its Vercel Agent, moving beyond alert triage and PR review. The agent now lives in the user dashboard, where it can investigate production issues, answer project questions, and take action upon approval. Because it’s built into the Vercel platform, it operates within security guardrails. Developers can invoke the agent directly from the project management dashboard.
Bun Rewritten in Rust: Performance and Stability Gains Detailed
Jarred Sumner published a detailed blog post on the complete rewrite of the Bun JavaScript runtime from Zig to Rust. The months-long effort aimed to improve performance, reduce memory footprint, and enhance cross-platform compatibility. The post provides a detailed account of the engineering decisions, memory management, and threading model changes.
Akashic with MemAttention Cuts LLM Inference Overhead for Long Contexts
Akashic introduces MemAttention, a low-overhead inference mechanism for LLM serving with long contexts. It targets the high prefill costs from accumulating context in multi-turn dialogues, tool calls, and cross-session workflows. By optimizing attention computation, MemAttention reduces redundancy, making LLM agents more efficient in continuous interaction scenarios, cutting inference latency and resource usage.
ResonatorLM Replaces Self-Attention with Resonant Field Mixing for Long Contexts
ResonatorLM introduces Causal Resonant Field Mixing, a replacement for self-attention in Transformers for efficient long-context modeling. Inspired by physical resonance, it mixes information via field interactions, maintaining parallel training efficiency while reducing memory and compute for long sequences. Early results show competitive modeling with lower inference latency.
NapMem: LLM Learns to Navigate Memory as Action Space for Personalized Dialog
NapMem transforms long-term user memory from passive retrieval to an active, structured action space for LLMs. Unlike traditional systems where models consume pre-selected evidence, NapMem lets agents navigate, combine, and manipulate memory fragments autonomously. This enables more accurate personalization in conversational AI, with superior results on long-term personalization tasks.
Benchmarking KV-Cache Optimizations Shows Tradeoffs Between Quality and System Performance
This work benchmarks KV-Cache compression techniques across models, tasks, budgets, and serving setups. It reveals clear tradeoffs between task quality (e.g., long-context accuracy) and system performance (throughput, latency), which prior studies couldn’t compare due to differing evaluation settings. The benchmark provides a unified reference for choosing the right compression strategy.
Learning to Control LLM Agent Harnesses with Offline Reinforcement Learning
This research proposes treating the LLM agent execution harness as a learnable control layer, not fixed infrastructure. Instead of just tweaking prompts or models, it uses offline reinforcement learning to optimize how the agent invokes tools, manages state, and executes steps. Results show that learning to control the harness yields greater agent performance gains than optimizing the LLM alone.
GitHub Agentic Workflows Automate Cross-Repo Documentation Updates
The GitHub Blog describes how the Aspire team uses Agentic Workflows to automate cross-repo documentation. When product code changes are merged, the system auto-generates corresponding documentation updates and creates PRs for SME review. This eliminates the gap between releases and documentation, ensuring docs stay in sync with code.
GitHub Copilot Enables Zero DNS Config for Pages, Live in 14 Minutes
GitHub Copilot now enables zero DNS configuration for GitHub Pages, letting users go from an empty repository to a live custom domain with HTTPS in about 14 minutes. It automates DNS resolution and certificate deployment, eliminating manual edits. This lowers the technical barrier for non-ops developers to launch sites.
Why AI Infrastructure must evolve for Agent Experience
Latent Space re-interviews Akshat Bubna, CTO and co-founder of Modal, to explore why Agent Experience is now viable and what Modal has learned building its new agent cloud. Bubna argues that current AI infrastructure is designed for request-response patterns, while agents require more flexible orchestration and state management.
OpenAI Analysis Flags Reliability Issues in Coding Benchmark SWE-Bench Pro
OpenAI published an analysis revealing issues with SWE-Bench Pro, a popular coding benchmark for AI models. The report indicates reliability and accuracy concerns, suggesting the benchmark may not accurately reflect real-world coding performance and has reproducibility issues. This raises questions about the validity of current AI coding evaluation methods.
OpenAI Lays Out Principles for Government and National Security Partnerships
OpenAI published a blog post detailing its approach to government and national security partnerships. The framework is built on three principles: responsible AI use, democratic accountability, and public safety. The post aims to clarify the company’s boundaries and ethical guidelines for sensitive collaborations. More details and case studies are expected to follow.
OpenAI and Walton Foundation Launch AI Skills Jams for K-12 Educators
OpenAI Academy partners with the Walton Family Foundation to launch AI Skills Jams for K-12 educators. The hands-on program teaches practical AI skills for classroom use, covering basic concepts, applications, and ethics. It aims to equip teachers with actionable methods to integrate AI tools into teaching, not just theory.
Kenton Varda Bans AI-Written Change Descriptions, Calls Them 'Worse Than Useless'
Kenton Varda declared a moratorium on AI-written change descriptions (PRs, commits, issues) for his team. He found AI-generated summaries focused on trivial code details, making them worse than useless for code review. The decision highlights growing concerns about AI’s effectiveness in producing meaningful developer documentation.
GitHub June Availability Report: 6 Incidents Caused Degraded Performance
GitHub published its June 2026 availability report, documenting 6 incidents that caused degraded performance across its services. The report includes the scope, duration, and remediation measures for each incident. Full details are available on the GitHub Blog.