Skills as procedural memory: teach your agent once, run it a hundred times

There is a category of feature that everyone calls "automation" but very few products actually deliver. The category looks like this: a user does some recurring task, the system notices the recurrence, the system offers to do it for them next time. In ten years of building this on different platforms (recipes in IFTTT, zaps in Zapier, automations in iOS Shortcuts) the experience has stayed the same: the user has to set the automation up explicitly, define the trigger, define the action, click through eight panels. By the time the automation exists, the user has already lost the joy of "the system noticed."

A skill is the actually-good version of that pattern. The user does the task once, in plain English. The agent does it. The agent recognises the shape of the request. The next time the user asks the same shape, the agent reuses what it learned. There is no setup screen, no trigger panel, no explicit "save this as a skill" step (although you can do that if you want).

This post is a tour of how skills work in Ajeris, what they are good for, what their failure modes are, and how to teach a skill on day one without thinking about it.

Procedural vs episodic

The other big memory layer is episodic memory: what was said and done. Episodic is "you booked a haircut at Felipe's on Saturday at 11." That is a specific fact about your week. Procedural memory is the recipe: "when I ask for a haircut, you book it at Felipe's, you check my calendar for an open Saturday morning, you confirm before charging." That is the how, separated from the what.

Most agent products only have episodic memory. They remember that you booked the haircut. They do not encode the recipe. So the next time you ask, they re-derive the recipe from scratch (and may get it wrong). Procedural memory is the missing layer.

Three kinds of skills, with examples

The taxonomy is loose, but a useful split is by who triggers the skill.

User-triggered skills. You ask the agent to do a thing in a specific way. The first time, the agent does the work. The second time, you give it less context because the recipe is now in its head.

You: book me a haircut
Ajeris: I have you at Felipe's last 3 times. Saturday morning is
        usually the slot you take. 9:30 or 11am this Saturday?
You: 11
Ajeris: Booked. 11am Saturday at Felipe's. Stripe deposit?

The "I have you at Felipe's last 3 times, Saturday morning is usually the slot" is the skill firing. The agent did not invent that pattern; it consolidated it from the last few episodic notes.

Time-triggered skills. A scheduled task with the agent's voice. "Every weekday at 7 a.m., send me a morning brief." The skill defines what a morning brief contains for you specifically (and you can tune it: "skip newsletters," "always include the weather"). Each tuning becomes part of the skill.

Event-triggered skills. A skill that fires when something happens. "Whenever a client signs a contract in HelloSign, send me a Slack message and add a row to the projects sheet." The first time, you ask the agent to do it once. The second time, the skill already knows.

How skills get saved

The lazy version is "save when patterns repeat." The agent watches for episodic patterns (three haircuts at Felipe's in three months) and proposes a skill. You either accept ("yes, that is how I want this to work going forward") or correct ("no, that was a fluke, I want it different next time").

The explicit version is "save this." You ask the agent to do something a specific way, and at the end you say "save that as a skill." The agent records the recipe with a name and a description. Next time you ask, the skill is already loaded.

The way most users actually end up using it is the third version: by accident. After a month of texting the agent, you notice it is doing things in your specific way without being asked. That is the skill layer working without the user thinking about it.

What a skill looks like under the hood

A skill is a small JSON record with a name, a description (in natural language), the recipe (also natural language), the channels it applies on, the version (skills auto-version when edited), and a timesUsed counter. Nothing fancy. The interesting thing is what the agent does with it: when a new request comes in, the agent searches its skill catalogue, finds the matching recipe, and uses it as a contextual hint when planning the response.

You can read your skills directly. Ask "what skills have you learned?" and the agent returns the list with titles and descriptions. Ask "show me the haircut-booking skill" and the agent shows you the recipe. Ask "patch the haircut-booking skill: always check Friday afternoon as a fallback if Saturday is booked" and the agent updates the recipe; the next invocation has the new behaviour.

Failure modes

Two real ones.

The agent over-fits. It learns "I always book at Felipe's on Saturday" from three booked haircuts and applies it the day you happen to be in another city. The fix is to ask the agent to reconsider: "I am in New York this weekend, do not use the Felipe's recipe." That correction itself becomes a hint the agent factors into future skills.

The agent under-applies. It does the task "from scratch" instead of using the skill it already has. This usually means the trigger phrasing was different enough that the agent did not match. The fix is to ask: "use the haircut-booking skill." After the first match the agent broadens its recognition automatically.

Both failure modes share a property: they are recoverable in plain English in 10 seconds. There is no admin panel. The skill catalogue is a conversation away.

What to skill on day one

A short starter list:

• "Send me a morning brief every weekday at 7am on WhatsApp."
• "When I ask for a meeting recap, always include action items split by owner."
• "When I ask to book a meeting, always offer two options 30 minutes apart."
• "When I ask for a restaurant, factor in [my dietary preference] without me mentioning it."
• "When I get a payment over [threshold], notify me on SMS, not WhatsApp."

Each of these takes one sentence. Each becomes a permanent how-to. By the end of the first week, the agent is doing things in your specific way without being asked.

Skills you do not have to write

The agent ships with a small set of built-in skills (booking, recap, push-to-phone, multi-channel handoff) that work out of the box. You can override any of them by saving your own version. The built-ins are documented in the docs for the curious, but most users never look at them.

Try it

Tell the agent one specific thing about how you like a small repeating task done. Watch it remember a week later. The first time you forget you trained it and the agent does the thing right anyway, you understand what skills are.