Hawkfry
All case studies
AIAI AgentsOperations

AI Employees With Long-Term Memory: How We Built Oplets

We treat our AI agents like members of the team. Each one has a job, a manager and a memory that grows over time, and they deliver real work alongside us.

AI Employees With Long-Term Memory: How We Built Oplets

Oplets are the AI agents we treat as employees. Each one has a job description, a manager who checks its work and a memory that builds up the longer it works with us. They work alongside our team on real delivery: picking up tasks overnight, watching a client's application for problems and reviewing code, with every piece of work handed back to a person.

What an AI Employee Is

An AI employee is an AI agent given a defined job, the access that job needs and a person it answers to. It differs from a chat assistant in two ways that matter. It keeps working when nobody is asking it anything, and it remembers what it has done.

We call ours Oplets. Our definition of an Oplet is an AI employee with a job description and a manager, and we run them much as we would run a new hire. We decide what the job is, give them the tools for it, check their work, and widen what they are trusted with as they prove themselves.

A typical AI assistant

  • Waits to be asked
  • Starts every conversation from nothing
  • Works only while someone is at the keyboard
  • Guesses when it is unsure
  • Hands back an answer

An Oplet

  • Has a job description and picks up its own work
  • Remembers its past work and learns how we do things
  • Works overnight and while the team is away
  • Asks its manager when a decision is not its to make
  • Hands back finished work for a person to review

Working Alongside Our Team

Our agents work on the same task boards as the rest of us and on the same client work. These are the jobs they do today.

What our agents are working on

  • Delivery working through labelled tasks overnight and opening code changes for review
  • Monitoring checking a live application for new errors every 30 minutes
  • Second review reviewing code changes that someone else has already merged
  • Support checking past answers before drafting a reply to a recurring question
  • Chief of staff drafting replies and preparing decisions for our founder

Work reaches an agent in three ways. Quick requests come in over chat. Longer pieces of work are labelled in our task tracker, and the agent picks them up overnight, writes its progress back to the task and opens any code change for review. It also checks a live application for new errors every 30 minutes, raises a ticket for each one, and proposes a fix when the bug is clear.

Each overnight run is limited to 30 minutes of work. We did that on purpose, so the agent cannot get stuck going round in circles and has to decide afresh whether a task is worth continuing.

How work reaches the agent

  1. 01ChatAny timeA quick request that needs an answer or a decision.
  2. 02Labelled taskOvernightAnything in the task tracker with the agent's label is picked up, in runs capped at 30 minutes.
  3. 03Error checkEvery 30 minutesA new error in a live application becomes a ticket, and a clear bug becomes a proposed fix.

Handed back for a person to review

  • A draft
  • An update on the task
  • A proposed code change

Never without a person's approval

  • Merging code
  • Changing a production database
  • Sending an email
  • Publishing anything sensitive

Like any new member of the team, an agent has things it must ask about first. It cannot merge code, change a production database, send an email or publish anything sensitive without a person's explicit approval. When it replies to an email, the reply is saved as a draft for someone to check and send.

Long-Term Memory

Many AI assistants start every conversation from nothing, so you spend the first few minutes of each one explaining your business again. Our agents remember. They keep a record of past conversations and work they can search, they read our company brain for what Hawkfry knows, and they check a set of notes on questions they are often asked, each linked back to where the answer came from.

They also learn how we like things done. When an agent works out a better way to handle a recurring job, it writes that down as a skill it can use next time.

That memory is only useful if it survives. Our agents are updated regularly, and an update must never wipe what an agent has learned, so we treat the three parts of a running agent differently.

  • Set by us

    Its role and personality

    Who the agent is, what its job is and how it should behave.

    Never overwritten by an update

  • Owned by the agent

    What it has learned

    Its memories and the skills it has written for itself.

    Never touched by an update

  • Generated from our templates

    Its settings

    Which tools it connects to, who may talk to it and which skills it may use.

    Replaced on every update

Before any update, we compare the running agent with the version we expect it to be. Anything useful it has built for itself is kept and added to the standard version for the next agent we set up, and anything nobody has reviewed is caught before the update goes ahead.

What Our AI Employees Have Delivered

Our agents work directly with our clients' employees, on the same delivery boards as the rest of the team. On one client's board the work ranges from fixes for errors in the live application to new features and reporting changes. Our consultant on the account estimated the agent was doing about six hours of coding work a night, ready for review in the morning.

The agent does not do this work alone. Every code change it proposes is reviewed by a person, corrected where needed and merged by them.

When our lead consultant was away, the agent covered monitoring on that client's application. It checked for problems every half hour and reviewed changes that the client's engineer had flagged for it, with a named person covering the alerts from other systems. The engineer told us afterwards that it had been very useful.

Our agents work through our own task board in the same way.

Why We Built Our Own Platform

An AI assistant on a laptop only works while someone is sitting at it. We wanted agents that kept going overnight and while we were on leave, without handing an AI unsupervised access to our code, our clients' data and our passwords. That meant giving each agent a computer of its own in the cloud.

Our first agent was set up by hand. It worked, but an early outage showed us how hard a hand-built agent is to recover, and how quickly it drifts away from what anyone thinks is running. A colleague had built a similar setup of their own at the same time. Two one-off setups were never going to be something we could run for anyone else, so we built one platform to run them all.

How We Keep AI Agents Secure

An agent that can act in your systems needs the same care as any other important system, and in places more.

  • A private machine for each client. Every client's agents run on a dedicated server, the industry's word for a computer in the cloud, and agents working for different organisations never share one. Agents on the same server share its storage, so we only put agents there that are meant to see each other's work.
  • Nothing reachable from the public internet. We reach each server over a private network, with a separate key for each person and no passwords.
  • Credentials kept in a secrets manager. They are handed to the agent when it is set up instead of being written into its settings.
  • A list of who may talk to each agent. Anyone not on the list is ignored.
  • Only the access the job needs. Agents get the skills we have approved for them, and database access limited to what their work requires.
  • Every version tested before it ships. Each new version is built, scanned for security problems and tested automatically before it can be released, and every server is backed up.

These rules are the same for every client. There is no setting that switches one off for a single client, so every agent is set up the same way by design.

Every update goes through the same steps, and we have written procedures for releases, updates, restoring from backup, and what to do when a memory or AI model provider goes down.

How an update reaches an agent

  1. 01Built and scannedEvery new version is built automatically and checked for known security problems.
  2. 02TestedIt is started and tested automatically before it can be released.
  3. 03ComparedThe running agent is checked against the version we expect, so nothing it learned is lost.
  4. 04Health-checkedThe update only counts as done once the agent is healthy and every chat channel is connected.
Our own agents get every update first. Anyone else gets it after it has run cleanly for a while.

What We Would Tell Anyone Considering One

Treat an AI agent the way you would treat a new hire. Give it a narrow job, a clear way for work to reach it, such as a label or an error alert, and a person who checks what it produces. Widen its responsibilities as it earns them.

Then plan for what happens around it. What decides whether you can trust an agent is what it can reach, who checks its work and how you recover when something breaks. If you plan to run one yourself, expect to update, monitor and recover it like any other important system. That is more work than most small teams expect.

AI Employee FAQs

What is an AI employee?

It is an AI agent with a defined job, the access that job needs and a person it answers to. It keeps working when nobody is asking it anything, and it remembers what it has done.

Do AI agents remember what they have done before?

Ours do. They can search their past conversations and work, read our company knowledge base and keep the skills they write for themselves, and our updates are built so that none of that is lost.

Does the agent make changes without a person checking?

No. Our agents hand back drafts, task updates and proposed code changes. A person reviews them, and merging code, changing a production database, sending email and publishing sensitive material all need a person's explicit approval.

Is it safe to give an AI agent access to our systems?

It can be, with limits. Give the agent only the access its job needs, keep it somewhere nobody can reach from the internet, and make sure a person approves anything that cannot easily be undone.

Which AI models do Oplets use?

Our agents run on large language models from more than one provider. That means an agent is not tied to a single AI vendor and can be moved if a provider has problems.

How is an AI employee different from an AI chatbot?

A chatbot answers when you ask it something. An AI employee keeps working when you are not there: it picks up tasks, watches for problems and remembers what it has done before. Our AI chief of staff case study shows what that looks like day to day.

Thinking About AI Employees for Your Business?

Oplets runs Hawkfry's own agents today, and early access for other businesses is opening soon. Join the Oplets waitlist to hear when it opens.

If you want to work out where an agent would actually help in your business, and what it would need to do it safely, talk to us.

Ready to put applied AI to work?

Tell us about your business and we'll map the fastest path to measurable results.