Programmer Mikhail Singatullin Files Offline Memory System for AI Agents With Russia's Software Registry

One developer, 22 years of experience, one application to the Russian software registry. Mikhail Singatullin, who has worked in systems integration, communication network design and server architecture, has applied to list a software package with long-term memory for AI agents. The stated use case is specific: agents in closed networks and under internet restrictions, where they suffer from "amnesia." CNews reported the news.

The problem: an agent that starts from zero every morning

A language model remembers nothing on its own. Everything it "knows" about your task sits in the context window, which is finite and resets with each new session. An agent that figured out your repository structure or your contract approval procedure yesterday asks about the same things again today.

In the cloud, external services plug this gap: vector databases, memory APIs, managed embedding stores. A closed network has no such services. The network is isolated, external APIs are unreachable, and security policy prohibits any data transfer outside. Unless memory is built into the package itself and runs locally, an agent in such an environment remains a "five-second goldfish."

This is the niche the author has targeted. Internet restrictions are no longer exotic for Russian companies: some organizations have segment isolation mandated by regulators.

What is known about the project

There is little public data so far. This is all that can be stated:

• Author: Mikhail Singatullin, 22 years in systems integration, communication network design and server architecture.
• Status: an application to the Russian software registry has been filed; no decision has been made.
• Purpose: long-term memory for AI agents.
• Environment: closed networks and restricted-internet conditions, meaning offline operation.
• Not disclosed: storage architecture, compatibility with specific models, benchmarks, license, price, whether any pilots exist.

The point of the registry is clear. For government customers and critical infrastructure operators, Russian origin of software is becoming a procurement requirement, so a registry entry opens this segment to a product. But filing an application is not the same as being listed: the package has to pass expert review, and the timeline depends on how complete the documents are. For now this is an intention, not a product with confirmed status.

A word on scale is also in order. Dozens of foreign and open-source projects tackle long-term memory for agents, but almost all of them are designed for cloud or hybrid environments. A solo author with an engineering background in networks and servers has chosen the infrastructure side rather than model research. That is a reasonable choice for offline environments, where what matters is not answer quality on a benchmark but whether the system deploys at all and runs inside the perimeter.

What to check before adoption

If you are responsible for AI in an organization with a closed network, the claim "we now have memory for agents" deserves a point-by-point review.

First: where the memories are physically stored and in what form. An agent's memory accumulates fragments of internal documents, correspondence and code, so it becomes a new body of sensitive data with its own encryption and audit requirements.

Second: access control. If several departments use the agent, you need to know whether it will "recall" for an accountant what it was told by a lawyer.

Third: a mechanism for forgetting and editing. Outdated and erroneous entries must be deletable, or the memory becomes a source of confident but wrong answers.

Fourth: which local models and embeddings are supported and how many resources it needs. In an isolated segment, GPUs are usually scarce, and the memory system must not consume them.

Until the author publishes a technical description and test results, it is sensible to treat the application as a signal of demand rather than a finished solution. The signal itself is telling: the need for offline memory for agents is now being addressed not only by large vendors but also by individual developers.