AI agents are turning past work into test beds and reusable knowledge

As AI agents take on more complex work, researchers are exploring ways to turn past tasks into reusable knowledge, procedures and testing environments for other agents

AI agents are moving beyond answering questions and generating content to carrying out multi-step tasks on behalf of employees. As they do more of this work, they are also generating detailed records of the decisions they make, the approaches they try, the tools they use and the outcomes they produce.

 

For now, much of that information is treated as an operational record. But researchers at Google and Microsoft are exploring ways to make it useful beyond the task that generated it. Past agent activity can be organised into reusable knowledge, procedures or even environments in which new strategies can be tested without repeating the original work.

 

 

Google's Dream-RSI (recursive self-improvement) research provides one example. It uses records of previous discovery processes as a replayable environment, allowing new exploration strategies to be tested against past results rather than requiring the agent to repeat the same work.

Microsoft on the other hand is working on systems that convert interaction histories into reusable knowledge, while Microsoft’s agent infrastructure is adding what it calls procedural memory, allowing agents to retain successful ways of carrying out tasks.

 

The developments point to a broader shift in how companies could treat AI-generated work history.

This is not AI training or simply AI memory

The idea described here is related to recursive self-improvement (RSI), but it is not about an AI system creating or training a better AI model. In conventional discussions of RSI, the system improves its own capabilities by developing or modifying models, algorithms or other parts of its AI stack.

 

Here, improvement happens through experience. An agent's previous work is recorded and reused to change how it approaches future tasks. The underlying AI model does not necessarily change. Instead, the system becomes better at a task by learning from what happened during earlier attempts.

 

It is also different from AI memory, although the two are closely connected. AI memory generally refers to a system retaining information from earlier interactions and retrieving it later to provide continuity or influence future behaviour.

The distinction here is what happens to the information after the task is completed. An agent may already use information gathered earlier in the same task to decide what to do next. For example, if an agent tries one approach, sees that it fails and changes its next attempt, it is using experience within that task. That does not necessarily create a reusable record for future tasks.

 

The emerging approach goes a step further. The experience from one completed task is captured, processed and made available to agents working on later tasks.

Past work can become a resource for future agents

Google's Dream-RSI research offers a particularly clear example of what this could look like.

 

The system is designed for AI agents that repeatedly explore possible solutions to difficult problems. Such systems can run through thousands of attempts, with each attempt producing information about what worked and what did not.

 

Instead of treating those results as a record to read later, Dream-RSI turns the history of an earlier discovery process into what the researchers call a "replay simulator". The historical record contains the decisions made during exploration and the outcomes of those decisions. A new strategy can then be tested against that recorded history without running the underlying agent again.

 

Essentially, this allows the system to use the structure of previous work to evaluate what might be done differently next time.

 

In the experiments, Dream-RSI recorded discovery trees generated during online exploration, converted them into a reusable simulator pool and then used alternative exploration strategies against those records before deploying a selected strategy for another round of exploration.

 

Google said that this approach reduced discovery-agent calls by up to 162 times against one existing system while maintaining or improving discovery quality in several tested settings.

 

The immediate objective is reducing the cost of experimentation. But the underlying idea has a broader implication: work performed by an agent can produce a resource that another agent can use later.

Microsoft wants agents to remember how work gets done

Microsoft Research is approaching the problem from another direction.

 

Its PlugMem research starts with a problem that sounds familiar to anyone who has tried to give an AI system a long history of previous interactions. More memory does not necessarily make an agent more useful. As interaction records grow, they can become filled with irrelevant information, making it harder for an agent to find what matters.

 

PlugMem therefore attempts to turn raw interactions into structured knowledge rather than simply storing old conversations.

 

The system extracts two broad types of information: facts and reusable skills. A fact might describe something learned during an interaction, while a skill captures something about how a task should be performed. These are organised into a structured memory system so that relevant knowledge can be retrieved for a later task.

 

That distinction could become important as agents take on recurring business processes.

 

Consider an agent handling customer service requests. A history of thousands of previous interactions could contain individual answers, but it could also reveal which checks need to happen before a refund is issued, which tools should be used and which mistakes should be avoided. The latter information can potentially be reused across future cases.

 

Microsoft is taking this idea further with procedural memory in its Foundry Agent Service.

 

According to Microsoft, procedural memory is intended to help agents retain successful execution patterns rather than simply facts. Agent activity is analysed to identify successful patterns, inefficient routes and missing steps. Those findings are converted into procedures describing when a particular approach should be used and what actions or checks should follow.

 

Microsoft reported about a 5 per cent improvement in its evaluations on STATE-Bench and Tau-Bench when procedural memory was enabled.

The amount of agent experience is also growing

This matters because the volume of work being delegated to agents is increasing.

 

OpenAI's August 2026 Enterprise Signals report describes a shift from AI assistance towards delegated execution. The report stated that as of June, Codex accounted for 64 per cent of combined Codex and ChatGPT output tokens among enterprise customers. It also found that weekly active enterprise Codex users had grown sharply since February in areas beyond engineering, including legal, sales, recruiting and marketing.

 

That expansion changes the potential scale of the problem.

 

An organisation using agents only for occasional questions may accumulate little useful operational history. An organisation using agents for software development, research, sales operations, recruitment and other recurring tasks could generate a much larger record of machine-performed work.

 

That history could contain successful workflows, failed approaches, tool choices, research paths and procedures that have emerged from repeated execution.

 

It is different from conventional corporate data because much of it is created as a by-product of getting work done.

The value may depend on how much of that history can be reused

This raises a potential change in how enterprises think about AI. Today, discussions around AI advantage often centre on access to models, computing resources, proprietary data and the quality of an organisation's prompts or workflows. Persistent agent experience introduces another variable.

 

Two companies could use the same underlying AI model while having different repositories of accumulated experience. One agent could have access to years of structured information about how a company's systems work, which approaches have failed and which procedures have previously succeeded.

 

That does not mean agent history is already a defensible competitive advantage. The research available so far shows ways of storing and reusing experience, rather than establishing that such history has become a measurable asset.

The more valuable the memory, the bigger the security risk

There is a catch to making agent history more valuable.

 

The more influence stored memory has over future behaviour, the more important it becomes to ensure that the information entering that memory can be trusted.

 

A July 2026 study called Bad Memory tested Anthropic’s Claude Code and OpenAI’s Codex across four models and found that malicious instructions already planted in persistent memory files could influence current and future sessions. The researchers found that getting agents to write malicious content into their own memory using untrusted external content was difficult, but once a malicious payload was present, it could persist and affect later tasks.

 

Microsoft has highlighted a similar risk, noting that persistent memory can allow an attacker to influence an agent over time rather than through a single prompt. It has described scenarios in which malicious instructions stored during one interaction trigger actions in a later, unrelated interaction.

 

That makes the quality of stored experience as important as its quantity.