Shlok Khemani

Shlok Khemani is a Mumbai-based AI researcher, programmer, and writer investigating how personal assistants remember their users and carry out work across conversations. His open-source projects, including OpenPoke, Claude Memory Tools, Rabbithole, and Vajra, translate research into AI memory and agent architectures into practical software.

Earlier in his career, Khemani worked at Decentralised.co, writing about cryptocurrency and developing products for that industry. His interests subsequently shifted from blockchain and decentralized computing toward language-model products, personalization, and autonomous agents.

In September 2025, he analyzed how ChatGPT builds persistent user profiles and how Claude retrieves earlier conversations, identifying distinct consequences for user control, inaccurate assumptions, stale information, and operating costs. Later that month, he introduced OpenPoke, a locally runnable reconstruction of the Poke assistant’s multi-agent architecture. Its design combines a coordinating conversational agent, persistent task agents, email, reminders, and multiple layers of memory. Khemani subsequently had a brief contracting engagement with Poke’s team.

In October, he examined Anthropic’s file-based memory tools and released Claude Memory Tools, demonstrating persistent, inspectable application memory in Next.js and Python. His later projects expanded beyond conversational assistants: Vajra, developed with collaborator Shruthi, moves Linear issues through planning, implementation, review, and pull-request creation, while Rabbithole turns documents into an explorable learning canvas connected to coding agents.

• Memory is product architecture. User profiles, searchable conversations, editable files, task histories, and external records serve different purposes. Khemani argues that serious teams should develop memory alongside their products instead of treating personalization as interchangeable retrieval infrastructure.

• Personalization has compute economics. Maintaining a profile consumes processing resources; including it in subsequent conversations adds recurring serving costs. Long profiles refreshed occasionally and shorter profiles updated frequently represent different compromises among accuracy, freshness, latency, and expense.

• Learning can happen outside model weights. Repeatedly synthesizing conversations into user profiles and applying those profiles to future interactions creates a practical learning loop, even without retraining an individualized model.

• Assistants need connected, correctable context. Separate AI products often construct incompatible portraits of the same person while overlooking evidence in inboxes and calendars. Khemani emphasizes visible, editable memory, recognition of contradictions, and human oversight when agents take action.