Skip to content

Kicker directives and memory

Kicker, AlertKick’s SRE agent, triages alerts, opens and summarises incidents, drafts after-action reports, and verifies changes. Two things make it behave like a colleague who has been on the team a while rather than a model seeing your fleet for the first time: directives you give it, and memory it keeps. Both are under Admin, then Agent Memory, and both need the SRE agent, which is included on Business and the trial.

A directive is an instruction in your words plus the deterministic effect it compiles to. The text is shown to the model as context on every matching decision; the compiled effect is enforced by AlertKick whether or not the model agrees. That split is deliberate: “always page the payments channel for anything on pay-1” is too important to leave to a prompt.

PartExample
Text”pay-1 and pay-2 are the card processors; treat anything on them as urgent and tell #payments-oncall”
ScopeThe whole workspace, one host, a host group (hosts sharing a security profile), or one rule
Compiled effectNotify immediately; a severity floor of high; a channel override of slack:#payments-oncall

Compiled effects available: notify immediately, severity floor, channel override (a Slack channel or another channel), audience (a specific user), and suppress. A directive can carry an expiry - none, 24 hours, or 7 days - so “ignore disk alerts on build-3 while we rebuild it” switches itself off.

Directives come from three places: the Agent Memory page (active as soon as you save), a chat with Kicker in the dashboard, or a Slack thread. Directives Kicker proposes from a conversation arrive as pending confirmation and do nothing until you confirm them on the page; the model never grants itself standing instructions. Each directive shows how many times it has matched and when.

Kicker writes to its memory as it works:

KindWritten when
EpisodeA triage concluded, an incident was summarised, a change was verified - what happened and what was decided
DigestA periodic roll-up of episodes for a host or the workspace
Host noteA fact about one host, from Kicker or from you
CorrectionYou told Kicker a conclusion was wrong

The memory browser lists entries by kind and scope with what they refer to (the alert, incident, or change), and a memory card per host and for the workspace summarises what Kicker currently believes. Delete any entry; a deleted memory is superseded rather than silently forgotten, so the audit trail stays intact.

You can add to it. A host note (“this box runs a legacy cron that spikes CPU at 02:00; expected”) or a correction (“the login from 203.0.113.7 you called suspicious is our bastion”) is read on the next relevant decision. Corrections are the fastest way to stop a repeated false positive without writing an ignore rule.

Alongside memory, Kicker keeps a rolling 30-day baseline per rule and per host: how often each detection normally fires, at what hours, from which processes. Triage compares a new event against the baseline, which is why the same event can be benign on a build server and suspicious on a database host. Baselines are visible on the Agent Memory page and rebuild themselves; there is nothing to configure.

The Skills section lists the procedures Kicker can run during triage - gathering evidence, checking a service’s topology, reading a runbook - and lets you reset any to its default. Custom skills are edited here for workspaces that want to add their own.

Every action is counted hourly by kind and host - triages that escalated, deep analyses, diagnostics run, incidents opened, resolved, and summarised, reports drafted, changes verified - with sampled examples, so the AI usage on the billing page and the activity view on the Agent Memory page reconcile with what it actually did.