Dhravya Shah

Dhravya Shah is the founder and CEO of Supermemory, where he works on AI memory infrastructure and context engineering. His work concerns how information from past interactions reaches an agent’s limited context window: what gets saved, how changing facts are reconciled, and when retrieval is worth the delay. In September 2026, he narrowed Supermemory’s focus to its memory API, discontinuing two applications built around it.[2][3][7][5]

In his technical account of Supermemory’s OpenClaw plugin, Shah describes a vector-graph layer that updates knowledge about users, with hooks saving both extracted memories and raw conversation chunks in the background. Retaining the chunks addresses a practical problem: the system cannot know at ingestion which details will matter later. He also describes deliberate forgetting, with irrelevant information decaying over time rather than remaining indefinitely in static files.[6]

His June 2026 essay, “Memory on the harness level”, gives a more qualified account of tool-based and observational memory: “And explicit is not the weak option, it’s the right one in the right place.” A coding or research agent can spend time reading existing memory before making an informed update. For an interactive personal assistant, he favors an observer that consolidates information outside the main agent loop. That moves work away from the reply a user is waiting for, but still costs inference, adds operational complexity and can leave the profile temporarily behind the conversation.[7]

Shah treats saving and recall as separate design choices. Agent-driven recall lets a model search, read and search again, spending latency to obtain better context. Harness-driven recall retrieves information before the model runs, offering a predictable retrieval step but depending on the user’s message as a useful query. He recommends longer retrieval loops for coding and research tasks, while interactive assistants benefit from context prepared before they answer. One combination he favors is implicit saving paired with explicit recall through a filesystem the agent can navigate.[7]

He also advocates pushing compact static and dynamic profiles into the context window: enduring information about the user alongside current activities. His example is a request to “find me the best monitor.” A search for “monitor” might find nothing in the user’s history, while a profile that already tells the model the user codes all day can still inform the answer. For Shah, personalization requires deciding what the model should always know, rather than relying entirely on query matches or the most recent memories.[7]

In a September 10, 2026 announcement, Shah said Supermemory had discontinued its company brain and Nova personal knowledge management platform on September 9 and refunded everyone who had been charged. Its MCP and agent plugins would continue operating through the developer platform, with the company concentrating on the memory API.[5]

Shah explained the decision through focus, positioning and potential competition with customers. Adding applications had made Supermemory harder to define; he acknowledged, “I built it, basically because we could.” He doubted that the team was best positioned to build the company-brain vision and wanted to enable other builders instead. He also worried that customers using its memory engine might see its applications as competitors, while saying this had not been an explicit issue. Discontinuing them, he argued, would let Supermemory work more broadly with other agent developers.[5]

Before the memory API became his central business, Shah was building open-source projects while studying. In his personal career account, he describes growing up in Mumbai, attending Arizona State University, starting Supermemory as a side project and cofounding Devlabs to build a community. He later worked at Cloudflare on Workers AI and special projects, including developer marketing, before dropping out to pursue his work full time.[2]

A concrete earlier AI project was his February 2024 live lecture assistant. Shah said he built it because he sometimes lost attention in class and hesitated to ask questions aloud. The open-source app transcribed lectures and let him ask questions about them without interrupting. He credited reuse of Notty’s backend, generated the initial interface with v0, and used the T3 stack, Embedchain and Deepgram for transcription. The project illustrates his earlier use of AI tools to build around a specific problem he experienced as a student.[8]