As AI-powered coding tools flood the market, one serious weakness has emerged: By default, like most LLM chat sessions, they are ephemeral – as soon as you close a session and start a new one, the tool forgets everything you were working on.
Developers have worked around this by using coding tools and agents to save their state in Markdown and text files, but this solution is hacky at best.
AI code review startup Qudo believes it has a solution with the launch of the industry’s first intelligent rules system for AI governance – a framework that gives AI code reviewers persistent, organizational memory.
The new system, announced today as part of Qodo 2.1, replaces static, manually maintained rules files with an intelligent governance layer. It automatically generates rules from real code patterns and past review decisions, maintains consistent rule health, enforces standards at every code review, and measures real-world impact.
For Itamar Friedman, CEO and co-founder of Qudo, this release represents a significant moment not only for his company but for the entire AI development tools sector.
"I strongly believe that this announcement of ours is the most important announcement till date," Friedman said in an interview with VentureBeat.
‘Memento’ problem
To illustrate the limits of current AI coding tools, Friedman cites the 2000 Christopher Nolan film MementoIn which the protagonist suffers from short-term memory loss and has to tattoo notes on his body in order to remember important information.
"Every time you call them, it’s a machine that wakes up," Friedman said of today’s AI coding assistants. "So all it can do is write everything it does to a file, before sleeping and restarting."
This approach—saving the context to Markdown files like agent.md or napkin.md—has become a common solution among developers using tools like cloud code and cursors. But Friedman argues that this method fails at enterprise scale.
"Think about heavy-duty software where you now have, let’s say, 100,000 sticky notes," He said. "Some of them are sticky notes. Some of those are pretty big explanations. These are some of the stories. You wake up and you get a task. The first thing is that [the AI] What that’s doing is starting to look for the statistically correct memo…it’s much better than not having it. But it’s very random."
From stateless to stateful
According to Friedman, the evolution of AI development tools has followed a clear trajectory: from autocomplete (GitHub Copilot) to question-and-answer (ChatGPT) to agentic coding within IDEs (Cursor) to agentic capabilities everywhere (Cloud Code). But he argues that all of them remain fundamentally stateless.
"For software development to truly revolutionize how we do software development for real-world software, it needs to be a stateful machine," Friedman said.
He explained that the main challenge is that code quality is inherently subjective. Different organizations have different standards, and even teams within the same enterprise may deal with problems differently.
"To reach truly high levels of automation, you must be able to adapt to the specific needs of the enterprise," Friedman said. "You must be able to provide code in high quality. But quality is subjective."
Qudo’s answer is what Friedman described. "Memory that is created over a long period of time and is accessible to coding agents, and they can then check and verify that what they are actually doing is in accordance with the subjective requirements of the enterprise."
How Qodo’s rules system works
Qodo’s rules system establishes what the company calls a unified source of truth for organizational coding standards. The system includes several major components:
- Automatic rule search: A rule discovery agent generates standards from the codebase and pulls request responses, eliminating manual authoring of rule files.
-
Intelligent Maintenance: A rules expert agent continuously identifies conflicts, duplicates, and outdated standards to prevent what the company says "Rule decay."
-
Scalable Enforcement: The rules are automatically applied during pull request code review, with developers providing recommended fixes.
-
Real-World Analytics: Organizations can track adoption rates, violation trends, and improvement metrics to prove standards are being followed.
Friedman emphasized that this represents a fundamental change in the operation of AI code review tools. "This is the first time that an AI code review tool is moving from reactive to proactive," He said.
The system surfaces code patterns, best practices, and rules based on its own library, then presents them to technical leadership for approval. Once accepted, organizations receive statistics on rule adoption and violations across their entire codebase.
A strong connection between memory and agents
According to Friedman, what distinguishes Kyudo’s approach is how tightly the rule system is integrated with AI agents – as opposed to treating memory as an external resource that the AI must search.
"In Kyudo, this memory and the agent are very much connected, just as they are in our brain," Friedman said. "It has a lot of structure…where the different parts are well connected and not separated."
Friedman said Qiudo applies fine-tuning and reinforcement learning techniques to this integrated system, which he credits with helping the company achieve an 11% improvement in precision and recall compared to other platforms and successfully identify 580 defects in 100 real-world production PRs.
Friedman offered a prediction for the industry: "When you look ahead a year, it will be very clear that when we started in 2026, we were in stateless machines trying to hack how they interact with memory. And by the end of 2026 we will have a very coupled approach, and Qodo 2.1 is the first blueprint of how to do that."
Enterprise deployment and pricing
Qodo positions itself as an enterprise-first company, offering multiple deployment options. Organizations can deploy the system entirely within their own infrastructure via cloud premises or VPN, use a single-tenant SaaS option where Qodo hosts a separate instance, or opt for traditional self-service SaaS.
Rules and memory files can reside wherever the enterprise requires – on its own cloud infrastructure or hosted by Cuido – addressing the data governance concerns that enterprise customers commonly raise.
In terms of pricing, Qodo is maintaining its existing seat-based model with usage quotas. Currently, the company offers three pricing tiers: a free Developer plan for individuals with up to 30 PR reviews per month, a Team plan at $38 per user per month (with a 21% savings for annual billing) that includes 20 PRs monthly and 2,500 IDE/CLI credits per user, and a custom-priced Enterprise plan with contact pricing that supports multi-repo context awareness, on-premises, Adds features like deployment options, SSO, and priority support.
Friedman acknowledged the ongoing debate in the industry about whether seat-based pricing makes sense in the age of AI agents, but said the company plans to address the topic more extensively later this year.
"If you get more value, you pay more," Friedman said. "If you don’t, we’re all good."
Initial Customer Feedback
Morag Breen of HR technology company Hibob, who was an early user of the Rules system, reported positive results for Qodo in a press statement shared with VentureBeat ahead of the launch.
"Kyudo’s rule system didn’t just bring forth standards that we spread across various places; It executed them," Brin said. "The system continually fine-tunes how our teams actually review and write code, and we’re seeing stronger consistency, faster onboarding, and measurable improvements in review quality across all teams."
Founded in 2018, Qodo has raised $50 million from investors including TLV Partners, Vine Ventures, Susa Ventures and Square Peg, along with angel investors from OpenAI, Shopify and Snyk.
<a href