The most counterintuitive thing we learned about building AI Agents

We’ve discovered something counterintuitive while building long-horizon AI agents at NonBioS: tools and memory can actually make your agent dumber.

I know this sounds backwards. To be fair, you need tools and memory to make LLMs useful and create agents out of them. But here’s what we’ve learned—every new tool and every new memory that you add to an agent adds to the cognitive load of the agent and leaves less capacity to tackle the actual problem at hand. This problem could be debugging complex software, or doing an analytics root cause analysis.

So what’s the solution to build effective long horizon agents? We took two radical approaches at NonBioS: Strategic Forgetting and what we call Grounded Linux capability for tool usage. Let me explain each.

Strategic Forgetting: Less Memory, More Focus

Instead of making agents remember everything, we invented Strategic Forgetting—an algorithm that continuously prunes the agent’s memory to keep it sharp and focused.

Here’s how it works: We evaluate information retention based on several key factors. Relevance scoring—we assess how directly information contributes to the current objective versus being tangential noise. Temporal decay—information gets weighted by recency and frequency of use, rarely accessed context naturally fades. Retrievability—if data can be easily reconstructed from system state or documentation, it’s a candidate for pruning. And source priority—user-provided context gets higher retention weight than inferred or generated content.

The algorithm runs continuously during coding sessions, creating a dynamic “working memory” that stays lean and focused. Think of it like how you naturally filter out background conversations to focus on what matters.

But it’s a work in progress. NonBioS does lose focus eventually and has to reboot. This is exactly like humans getting tired by end of day and needing to sleep. But if you read a novel for breakfast, you might need to nap by noon.

Grounded Linux Capability: Rethinking Tool Calls

NonBioS does zero tool calls in the traditional sense. Even though NonBioS has access to all the tools of a Linux environment, it does so through what we call a Grounded Linux Capability. The grounded capability means that the NonBioS orchestration layer imputes tool calls through NonBioS responses which are grounded in the training set. Because the tool call doesn’t pollute the inference time context, it maintains better cognitive capacity over time.

Every user in NonBioS gets a full Linux VM with root access. You can ask NonBioS to pull in your source code and implement any feature. The context is all managed automatically through Strategic Forgetting.

Both Grounded Linux capability and state-of-the-art Strategic Forgetting are the critical components that allow NonBioS’s long horizon autonomy to manifest. Right now NonBioS is able to run multi-hour software engineering sessions with human-in-the-loop guided autonomy. However this capability is constantly being extended and we have no reason to believe that we’re hitting any limits yet. There’s little reason to believe that we won’t reach multi-month autonomous sessions by early next year.

What this could mean: NonBioS will be able to work on large scale software projects almost completely autonomously, checking in with you periodically to align expectations and make sure you’re aligned on the final outcome.

The challenge is getting the balance right—too aggressive with forgetting and you lose important threads, too conservative and you’re back to the goldfish problem - agent drowning in its own context. But we’ve tried it out in very complex code bases and it works remarkably well. Once you see how well it works, it becomes easier to imagine that the days of using IDEs to edit code might be numbered.

The core insight? At NonBioS we take the opposite approach to most agent systems. Instead of making agents remember everything, we make them forget what’s not important. This keeps them sharp and focused on what truly matters.