← All speakers

Sparsh Jain

Conference affiliation: CentralAgent AI

Sparsh Jain builds agent memory that carries lessons from one teammate’s session into another’s. As head of AI engineering at CentralAgent in July 2026, he described a problem that more conversation history had failed to solve: an agent could retain facts yet keep repeating mistakes its users had already corrected. His team’s response was to preserve reusable lessons, the circumstances in which they worked, and the history of decisions that later changed.

On this page

From specialized models to agent infrastructure

Sparsh’s earlier work included biomedical language-model research at Nanyang Technological University in Singapore. In 2022, he co-authored MedBERT, a six-author collaboration addressing the difficulty of transferring specialized medical models across different kinds of biomedical text. The team continued pretraining Bio ClinicalBERT on clinical challenge data, biomedical research corpora and filtered Wikipedia text, then evaluated named entity recognition across ten datasets. The paper lists Sparsh with a BITS Pilani affiliation and does not specify his individual contribution.

He subsequently worked on quantitative machine learning for risk at Wells Fargo, holding an assistant vice president role, and led AI engineering at Key AI, working on agentic security systems. By July 2026, he was leading AI engineering at CentralAgent.

CentralAgent captures company knowledge from meetings and agent sessions and makes it available to employees and agents through a shared Model Context Protocol endpoint. The useful knowledge includes why a service became the default, which incident fix worked, and what checks must precede a deployment. Its product examples distinguish reading a configuration from recovering the decision behind it: an agent might find the shipped read path in code, while company memory supplies the earlier latency incident, cost rationale and deliberate exceptions. These are examples of the team’s product, rather than accomplishments attributable to Sparsh alone.

Preserving lessons without accumulating every detail

In his account of CentralAgent’s memory work, Sparsh explains why supplying more history eventually became counterproductive. It increased cost and response time while reducing accuracy. The team moved toward selecting and revising experience rather than presenting every stored detail to the agent.

He distinguishes three levels of reusable experience:

  • Cases preserve an incident, its outcome and what the team learned.
  • Strategies capture an approach that worked under particular conditions.
  • Skills turn sufficiently general lessons into capabilities an agent can use repeatedly.

The distinctions preserve both a lesson and the circumstances that made it useful. Sparsh also argues for retaining changes of mind. A superseded judgment can become inactive without disappearing: routine questions retrieve active knowledge, while deeper investigation can recover the sequence that led to the current decision. That history helps an agent recognize an approach the team has already tried and abandoned.

Retrieval must still select the right lesson for the situation. Sparsh illustrates the problem with an agent that knows its user has homes in San Francisco and New York. Asked for directions home after work in San Francisco, it can retrieve a valid address and send the user across the country. Correct storage alone does not make the answer appropriate.

Testing what memory changes

Sparsh’s team connects memory updates to recurring evaluation. After sessions and at the end of the day, review loops extract lessons, revise priorities, remove unhelpful material and test proposed changes before applying them. He advocates using models as verifiers within these loops, allowing the system to assess updates instead of requiring a person to review every one.

He reported that context gathering took about twenty minutes in many earlier cases; when the agent had gathered the context before and knew what to retrieve, it took about two minutes. That comparison describes his team’s experience with repeated context gathering, rather than a general benchmark for agent performance.

Sparsh also leads the AI Tinkerers Hyderabad chapter.

1 conference talk

Loading more talks and highlights…

References