Rishabh Singh is betting the road to AGI runs through memory

Opinions expressed by Digital Journal contributors are their own.

In a survey conducted for a March 2025 report by the Association for the Advancement of Artificial Intelligence, 76% of 475 respondents said scaling up current AI approaches was unlikely or very unlikely to produce artificial general intelligence, a category of systems broadly understood to learn and reason across a wide range of tasks.

That skepticism parallels a question at the center of Rishabh Singh’s research. Singh, co-founder and chief technology officer of JoinMeNow Inc., also runs Helix Mind, an independent research effort he describes as focused on capabilities that scaling alone may not address, namely persistent memory, continual learning, and adaptation to unfamiliar problems.

The survey respondents were mostly academics, and about a fifth were students, according to the AAAI 2025 Presidential Panel report. They answered as major AI labs were investing heavily in larger models, training infrastructure, and computing capacity.

Singh argues that current language models still face limitations in persistent learning, updating internal knowledge without degrading earlier capabilities, and adapting efficiently to unfamiliar problems.

“The real breakthroughs I’m seeing come from systems that can remember, reason, and teach themselves,” Singh wrote in response to written questions.

The problem of keeping what you learn

Many AI products now remember users across conversations by storing notes outside the model and pulling them back in later. That memory lives in the product, not in the model’s weights, the internal parameters set during training. The model itself does not learn from the exchange.

Changing those weights brings its own problem. Updating a trained network on new data can degrade what it already knew, a problem highlighted in classic 1989 work by Michael McCloskey and Neal Cohen and now commonly called catastrophic forgetting. A third challenge is adaptation, inferring the rules of an unfamiliar problem from only a handful of examples.

Singh calls his proposed answer cognitive models, which he defines as “architectures that integrate memory, reasoning, and self-learning.” The phrase is his own framing of capabilities that have long been studied separately across AI research.

Through Helix Mind, Singh says he is exploring new language model architectures and training methods aimed at that combination. Helix Mind has not released papers, code, or benchmark results. Without them, the approach remains a research direction rather than a demonstrated result.

From trading systems to learning systems

Singh’s route into the problem ran through quantitative finance and production software rather than an academic AI lab. He says he spent years writing ultra-low-latency trading software in C++ for financial institutions, and later built machine learning models for credit risk and pricing.

That work, he says, forced him to think about efficiency, memory management, reliability, and whether a model holds up on data it has never seen. In his description, a trading engine executes rules people design ahead of time. It gets faster, but it does not learn between updates. Singh describes the goal of his research as building “AI that thinks instead of just reacting.”

JoinMeNow Inc. develops Join-Me-In, a location-based iPhone app for meeting people nearby, released on Apple’s App Store in May 2026. Singh says his work there centers on live location data, secure sign-in, and privacy. He also identifies himself as founder of Nova Infinity, which he describes as developing specialized AI products on shared infrastructure for AI, identity, and billing, and Nova Reach, a campaign automation business.

Singh argues that systems designed to learn while running would still face familiar production constraints, including latency, cost, reliability, monitoring, and failure recovery. He has written that the slow pace of research at Helix Mind keeps him honest “about what’s actually possible versus what’s just good marketing copy.”

A test that rewards learning

ARC-AGI tests a related problem. It asks whether an AI system can infer an unfamiliar rule from a handful of examples and apply it correctly to puzzles built from colored grids. The benchmark was created by François Chollet, who co-founded the ARC Prize competition with Mike Knoop. In controlled testing with more than 400 participants, every ARC-AGI-2 evaluation task was solved by at least two people in two attempts or fewer, according to the ARC-AGI-2 technical paper.

In the 2025 competition, 1,455 teams submitted 15,154 entries under fixed computing limits. The top competition entry scored 24.03% on the private ARC-AGI-2 evaluation, and the grand prize went unclaimed, ARC Prize reported. In his analysis of those results, Knoop argued that current AI reasoning performance “is tied to model knowledge.” He added, “Human reasoning capability is not bound to knowledge.”

None of this shows that Singh’s approach works. Neither he nor Helix Mind has reported ARC results. The benchmark shows why the problem he chose is hard, not whether his answer is right. Researchers also disagree on how to define AGI, how to measure it, whether it will arrive and, if it does, on what timeline.

Singh spent the early part of his career making software respond faster. His research now asks whether software can keep what it learns and use it the next time.