A research lab studying
cognitive architectures
for the next generation of AI.
Today's language models are stateless transformers. They have no memory and cannot reason across time.
We are building the next generation of AI that consolidates knowledge, the way biological intelligence does.
Why now,
why us.
Why now
The current generation of large language models has reached a clear plateau. Models continue to grow larger and context windows continue to grow longer, but the fundamental limitations remain. These systems do not remember. They do not reason across time. They do not consolidate knowledge into structured understanding. They do not differentiate what they know from what they infer.
These are not engineering details that scale will solve. They are architectural assumptions that need to be questioned.
At the same time, three things are converging. Cognitive science has produced sixty years of detailed theory about how biological memory actually works — episodic and semantic distinctions, complementary learning systems, consolidation during sleep, forgetting curves. Open-source language models have become powerful enough to serve as the substrate for new cognitive architectures. And the demand for AI systems that genuinely persist, reason, and remember is no longer hypothetical — it is being articulated by every serious user of these tools.
The gap between what current AI does and what cognition requires has never been more visible.
This is the moment to build what comes after the stateless transformer. Not by training larger models, but by constructing the cognitive architectures that should sit alongside them. That is the work of this lab.
Why us
Neural AI is built at the Indian Institute of Technology Madras — the same institution that has produced some of the most rigorous applied research in the country. We sit close to the academic foundations our work draws from and close to the engineering culture that turns research into working systems.
We are not trying to compete with frontier model labs on scale. We are taking a different bet: that the next generation of AI capability comes from architectural sophistication, not parameter count. That bet is best made by a small team that can move quickly, study carefully, and build deliberately.
Anant 1.0 is the first manifestation of this bet. A working cognitive memory system, currently in private alpha, that implements persistent structured memory, belief state architecture, multi-channel retrieval, and consolidation cycles informed by complementary learning systems theory. It is one system inside a longer research agenda.
We expect to be wrong about specifics. We expect to be right about the direction.
What we are not building
We are not training foundation models. We are not building agent frameworks. We are not adding memory as a feature on top of existing chat interfaces. Each of those is being done well by other organizations.
We are building cognitive architectures — the structured systems that determine how AI represents, retrieves, consolidates, and reasons over information. The model is a substrate. The architecture is the work.
Our research agenda.
Neural AI is pursuing a research agenda spanning nine areas of cognitive architecture.
Persistent Cognitive Memory
Belief State Architecture
Multi-Channel Retrieval Fusion
Memory Consolidation Cycles
Causal Emotional Intelligence
Continual Learning Architectures
Memory is not a downstream feature; it is the fundamental substrate of cognition. The current paradigm of stateless transformers necessitates that every session initiates from a zero-state, rendering them computationally amnesic. Attempting to mitigate this via infinitely scaling context windows is a brute-force patch, an unsustainable trajectory fraught with latency and quadratic compute costs. It represents a profound architectural failure that requires a first-principles resolution.
The trajectory toward general intelligence cannot rely solely on parameter scaling. It demands structured, stateful cognition: discrete episodic memory encoding, dynamically updated belief states, asynchronous memory consolidation, biologically-inspired forgetting curves, and rigorous temporal reasoning mechanisms.
These constructs are not novel inventions. They are rigorously established principles derived from six decades of cognitive science and neuroscience research—paradigms that the contemporary deep learning consensus has largely bypassed in favor of raw compute optimization.
Our research lab is systematically engineering the next generation of cognitive architectures, embedding true persistent state into the core inference loop. We are building systems that translate this academic foundation into deployable computational reality.
First System
Anant 1.0
Our first cognitive memory system.
Anant 1.0 is the first manifestation of our research, a working system that implements persistent cognitive memory, belief state tracking, multi-channel retrieval, and consolidation cycles.
Currently in private alpha with select partners. Public release information will be shared when we are ready.
Safety, built in
Latent Manifold Integrity
We constrain episodic embeddings with typed latent manifolds, uncertainty bounds, and representation audits so memory retrieval remains calibrated instead of merely semantically adjacent.
Belief Graph Topology
Belief states are represented as evolving causal graphs with attention-weighted edges, contradiction checks, and provenance links across long-horizon inference traces.
Consolidation Dynamics
Short-term interaction traces are compressed through replay, decay, and semantic distillation cycles, separating durable knowledge from transient prompt noise.
Efficiency Wanted
Engineer the future of intelligence.