OpenAI and Google both shipped models with context windows exceeding 10 million tokens this week, continuing the arms race to build 'machines that never forget.' GPT-5.4's extended context allows it to process entire codebases, legal document repositories, and multi-hour meeting transcripts in a single pass. Google's Gemini 3.1 Pro matches the milestone with a focus on long-context retrieval accuracy, publishing benchmarks showing 97% recall on the 'needle-in-a-haystack' test at 10M tokens.
The practical value remains debated. Early enterprise feedback indicates that while large context windows eliminate the need for RAG in many scenarios, retrieval accuracy degrades significantly past approximately 2 million tokens in production workloads. 'The context window is only as useful as the model's ability to find the right information in it,' noted one AI engineer at a Fortune 500 firm. Both labs are investing in 'context distillation' techniques that prioritize relevant information within the window rather than simply expanding it.
Perplexity launched its 'Personal Computer' product, a software platform that transforms a Mac Mini into an always-on AI agent with persistent memory and autonomous tool access. The system maintains state across sessions, learns user preferences over time, and can execute multi-step tasks — from managing email to conducting research — without requiring continuous human input. Perplexity positions it as a 'co-pilot that doesn't forget what you asked it to do yesterday.'
The product represents a new category of AI deployment: locally persistent agents that combine the memory benefits of cloud AI with the privacy guarantees of local execution. The platform uses a hybrid architecture where sensitive data remains on-device while compute-intensive tasks route to Perplexity's cloud. Pricing starts at $99/month for the software-only tier; a bundled Mac Mini option is available for $999.
Anthropic rolled out persistent memory for Claude Enterprise customers, enabling the AI to maintain context across sessions, recall user preferences, and reference past conversations. The feature addresses one of the most requested capabilities from enterprise users who found themselves repeatedly re-explaining context to what was effectively an amnesiac AI. Claude's memory is stored in encrypted per-user profiles and can be reviewed, edited, or deleted by users at any time.
The memory feature is implemented using a vector-based episodic memory system rather than simply extending the context window — a design choice Anthropic says makes memory both more efficient and more auditable. Each memory entry is timestamped, attributed, and searchable. The feature is available first on Claude Enterprise at $100/user/month, with Team and Max tiers planned for Q3. Anthropic published a transparency report alongside the launch, detailing what the system remembers and how users can audit it.