MediaNama published an interview on November 13, 2026, featuring Diwank Singh Tomer and Ishita Jindal, the co-founders of MemStore. This shared memory system for teams and their AI agents was developed by the founders of Julep AI. The discussion focused on how the platform organizes knowledge, handles conflicting information and manages data retention under privacy regulations.
Tomer and Jindal described the system’s architecture as comprising four distinct layers: episodic, implicit, prospective and semantic memory. Episodic memory records time-stamped events and sequences, while implicit memory tracks the beliefs and current knowledge states of individual participants. Prospective memory stores future tasks and goals, and semantic memory maintains static facts within a large knowledge graph. This structure allows the system to model information in various ways, with embeddings enabling both retrieval and clustering of related ideas across different team members' work.
A central challenge addressed in the report is the consistency of outputs. Tomer stated that while deterministic outcomes are impossible due to the variable nature of language models and context windows, consistent outputs are achievable through careful architecture. The system flags contradictions between records rather than silently resolving them. For instance, if one record indicates a meeting between two individuals and another contradicts this, the system maintains both until logical updates reconcile the discrepancy. This approach acknowledges that memory systems are inherently lossy and that models cannot predict all necessary context in advance.
Regarding data deletion, a critical issue under data protection laws, Tomer explained that the system does not simply remove nodes from the knowledge graph. Instead, when a user requests that a fact be forgotten, the system reduces the weight of connections to that fact to zero. This ensures the information never appears in queries without the technical complexity of deleting nodes. Jindal added that the system distinguishes between data worth remembering and operational noise, such as raw event analytics, focusing instead on how tools are used and who accesses them.
Scaling the system to handle years of institutional knowledge presents significant challenges. Tomer noted that managing long time horizons requires agents to spend tokens organizing and classifying information to determine relevance. The founders emphasized that non-agentic memory systems struggle with this scale due to noise and overlapping data. The interview highlights the ongoing development of these systems as they attempt to balance comprehensive data retention with efficient retrieval and privacy compliance.