AI & Automation Briefing - August 26, 2026
Anthropic Unifies Claude Chat and Cowork Memory
Anthropic announced on August 25 that Claude now shares a single memory system across its chat and Cowork environments. Previously, context built up in chat did not carry over to Cowork, forcing users to re-enter project details when switching to action-oriented tasks. With the unified system, anything Claude learns during a conversation is available immediately in Cowork, even mid-session, without waiting for the chat to end. Users can view, edit, or delete what Claude has stored. By default, Claude will not retain sensitive personal data such as health information, political views, or religious beliefs, though users can opt in to storing those categories. Government-issued ID numbers, Social Security numbers, and similar data are permanently excluded from memory. The feature is on by default for Free, Pro, and Max plan users across web, desktop, and mobile.
OpenAI's Jalapeño Chip Posts Benchmark Wins Over Nvidia Blackwell on Tokens and Power Efficiency
OpenAI presented the first benchmark results for its Jalapeño inference chip at the Hot Chips conference on August 25, 2026. Tested against the SemiAnalysis InferenceX benchmark, Jalapeño outperformed current state-of-the-art inference processors, including an Nvidia Blackwell system, on both tokens per user and throughput per kilowatt. The chip is designed to reduce bottlenecks during the prefill and communication phases of inference by keeping model state and KV cache local and coordinating compute, memory, and networking per inference phase. OpenAI developed Jalapeño with Broadcom and used its own AI models in the design process. Deployment is expected in small volumes by end of 2026, with broader rollout in 2027. OpenAI plans to treat Jalapeño as a multi-generational platform, co-developing AI products, models, chips, and memory together. For teams running high-volume agentic workflows, the performance-per-watt gains signal a shift toward cheaper, faster inference at scale.
AutoSaddler Automates LLM Agent Harness Improvement Using Failure Traces
Microsoft researchers published AutoSaddler, a framework that automatically optimizes LLM agent harnesses by learning from offline execution failures. Agent harnesses, the scaffolding that governs prompts, tool configurations, and control logic, currently require manual design and degrade on long-horizon tasks where small failures compound across many steps. AutoSaddler treats harness improvement as a code-level optimization problem, diagnosing failure traces in mini-batches, generating structured patches, and selecting validated updates iteratively. Tested on GAIA2, SWE-Bench Pro, and Terminal-Bench 2.0, the system delivered performance gains of 9.0, 9.6, and 10.0 percentage points respectively over base harnesses. For teams building agentic workflows in tools like n8n, the approach points toward self-correcting agent scaffolding that reduces the manual overhead of keeping multi-step automations reliable.