Corelayer

It’s 9:05 AM on Monday morning and it’s your first week in the on-call rotation. You should be able to debug simple issues. What about something more complex? You can try root-causing everything from first principles, but at best it’ll take longer and, in reality, you’ll end up pulling in someone more experienced.

An AI agent can fetch and summarize logs, query data, understand metrics, and inspect code. But without the ability to learn from experience, every day is its first day on the job. At Sevvy, we’ve built a custom memory layer that mimics the process that experienced engineers rely on: pattern matching against past issues, remembering what happened and why, and applying that to the current state of the system to perform a more accurate and efficient investigation.

What is an "Experienced" AI Agent?

Senior engineers don’t read logs in isolation. They develop instincts for systems they’ve worked on long enough, comparing symptoms against years of past issues. Sometimes this is also explicit, where tickets, incidents, and post-mortems are documented and referenced when similar patterns reappear.

We think the reasoning process of an “experienced” agent should look like that of an experienced human. Sevvy’s memory layer distills insights from each failure, curating them over time based on human feedback. These include the pattern that caused the issue, the assumptions that may have been violated, the signals that actually mattered during the investigation, and the type of fix that resolved the problem. This information is stored at a high enough level of abstraction to remain useful as the underlying implementation details change.

When a new error or anomaly arrives, Sevvy starts by asking the same question a human engineer would: “Have I seen something like this before?” If a similar pattern exists, that becomes the starting point for the root-cause hypothesis, allowing the agent to run a focused investigation rather than starting from scratch.

Technical Overview

Our memory system is built on a handful of simple concepts.

Feedback
We capture feedback from the engineer at the time of closing an issue, indicating things like the accuracy and usefulness of the agent’s investigation process, root cause, and suggested fixes.

Normalization
Each issue is distilled into a structured record that captures a summary of what went wrong, why it happened, how it was fixed, and human feedback on the quality of the agent’s analysis and fix. We normalize this by removing things like line numbers and timestamps.

Embedding
We generate and store an embedding of the issue using qwen3-embedding-4b. We then use this to search for semantically similar past issues when new issues appear.

Hypothesis-driven debugging
When a match is found for a new issue, the agent uses it as a starting point for its root-cause hypothesis. It then investigates to either confirm or reject (and change) this hypothesis. We prompt the main orchestration agent carefully so we don’t “overfit” to this hypothesis.

Continuous maintenance
As new issues come in and are closed with human feedback, we add, update, and delete memories using a memory-refining sub-agent.

Much of this design came from applying the mem0 architecture to our use case.

Impact on Issue Resolution

With each resolved incident, Sevvy develops a deeper understanding of how your system behaves. Future investigations can be short-circuited and resolved significantly faster. Incidents that once required multiple rounds of log parsing, code search, and probing of data now begin with informed guesses backed by historical context.

This does more than speed up resolution. It improves accuracy on more complex issues that it didn’t get right the first time around. It also gives your team a consistent source of operational knowledge that doesn’t disappear when someone leaves or switches teams.

Another common, practical implication of memory is that you can give the agent feedback on which issues to ignore in the future. Maybe there’s a known, recurring issue that’s already in your backlog but you haven’t prioritized. Sevvy won’t re-investigate future occurrences, reducing alert noise.

Over time, Sevvy starts operating like that senior engineer that’s been around forever. The one that you know you can trust and rely, especially for your most complex and critical issues.