Our research teams investigate the cognitive architectures, memory persistence, and belief state tracking of AI models—so that artificial intelligence has a positive impact as it becomes increasingly capable.
The mission of the Cognitive Memory team is to develop persistent memory layers that allow models to encode and recall episodic data indefinitely.
This team works to build world models within AI, ensuring systems can track internal belief states and update them based on new evidence.
Focusing on multi-channel data synthesis, this team explores how AI can retrieve and fuse information from disparate knowledge graphs in real-time.
Investigating the boundaries of persistent systems to ensure memory consolidation cycles remain aligned with human intent and safety constraints.
Research manifestation
Research / April 24, 2026
Anant is the working system where our research agenda becomes architecture: persistent memory, belief state, retrieval fusion, consolidation, and safety boundaries around what should be stored or forgotten.
Memory / May 2, 2026
We explore the mathematical foundations of encoding temporal events into a persistent cognitive layer.
Belief States / April 15, 2026
How persistent systems can maintain consistency across long-form interactions using dynamic belief updating.
Knowledge Graphs / March 20, 2026
Integrating graph-based retrieval into the core inference loop to solve for logical grounding.
Foundational research papers and systems work that inform Neural's agenda. These citations are not claims of authorship; they are the technical lineage behind our work on memory, belief state, retrieval, and cognitive architecture.
Packer, Fang, Patil, Lin, Wooders, and Gonzalez
A recent systems paper on managing memory beyond a fixed context window. It is relevant to Neural's research on structured cognitive memory, belief state, and consolidation.
Park, O'Brien, Cai, Morris, Liang, and Bernstein
A useful system reference for memory, reflection, and planning loops in interactive agents. Neural studies similar ingredients under stricter architecture, auditability, and user-state constraints.
Yao, Zhao, Yu, Du, Shafran, Narasimhan, and Cao
Shows how explicit reasoning traces can improve action selection. Neural's belief-state work asks how those traces can persist safely and be revised across long-horizon interaction.
Lewis, Perez, Piktus, Petroni, Karpukhin, Goyal, Kuttler, Lewis, Yih, Rocktaschel, Riedel, and Kiela
The modern reference point for combining parametric and non-parametric memory. Neural builds on this idea but focuses on persistent user-state, belief objects, and temporal memory rather than static document retrieval alone.
Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, and Polosukhin
The foundation for the model substrate Neural builds around. Our work treats the transformer as powerful but stateless, then studies the memory and state architecture that should sit beside it.
Sukhbaatar, Weston, Fergus, and others
A key reference for multi-hop reasoning over stored facts. Neural's retrieval fusion work extends this direction into persistent, multi-channel cognitive context construction.
Graves, Wayne, and Danihelka
An early demonstration that neural systems can learn to interact with external memory. It motivates the question of how modern language models should read, write, and control persistent memory.
McClelland, McNaughton, and O'Reilly
A central cognitive-science foundation for separating fast episodic storage from slower semantic consolidation. This is directly relevant to Neural's consolidation-cycle research.