Skip to content

Nightshift's memory

Nightshift remembers what it learned in your organisation. Each time it works on an alert or an incident, it keeps one short lesson: what the failure looked like, what caused it, what was done, and whether that held. The next time something similar fires, Nightshift is shown the most similar lessons, opens the ones that fit, and cites them like any other evidence.

So the second time shop/ledger-api is OOM-killed after a release, the investigation can say it has happened before, which fix held then, and which did not.

  • Your organisation’s. It is stored in your region’s Seylo database with the rest of your organisation’s data, and nowhere else.
  • Only your admins see it, on Nightshift → Memory. Everyone else sees Only admins can see the organisation's memory., and on an incident only what the run opened.
  • Never shared with another organisation, and never used to train a model. A lesson reaches a model only as evidence inside one of your own investigations, through the model your organisation chose.
  • No Secret value is ever kept. A field that carried a secret keeps its name only. Lessons hold no log lines, no evidence text and no email addresses: only the evidence’s numbers and kinds.
  • It adds nothing Nightshift may do. The list of actions Nightshift can propose, its targets, its namespaces and its autonomy are exactly the same with memory on or off.

Memory is on by default.

A lesson is written when an investigation ends with a finding that cites its evidence, and when a fix reaches its outcome: verified, did not verify, undone, rejected, no longer needed, or did not take. An investigation that could not determine a cause is not remembered, and neither is a refusal.

Part Content
Where Cluster, namespace and workload. Generated suffixes are dropped, so ledger-api-7d9f8b6c5-x2k4p is remembered as ledger-api
Alert Its reason, such as OOMKilled
Symptom The reason and the alert’s message, shortened
Cause The finding Nightshift reached
Fix What was done, in one line and at most four lines of its diff
Outcome The latest outcome, with a count of each, for example verified 3, undone 1
Evidence The numbers and kinds of the evidence the cause rested on - not its text
Skills The skills the run used
Model The model that wrote the finding

A lesson is at most 1,200 characters. The same workload, alert, cause and fix within 30 days is one lesson with its counts going up, not a new one each time.

When an investigation starts, Nightshift is shown up to 5 past lessons as a short index, most similar first:

  1. the same workload;
  2. the same namespace and alert;
  3. the same alert anywhere in your organisation.

Fixes that verified rank first. A fix that did not hold is still shown, labelled as tried before and not held, and Nightshift is told not to propose it again unless this run’s evidence shows why it would hold now. Recent and repeated lessons rank a little higher.

Nightshift opens at most 3 lessons per run. Each one it opens becomes one numbered piece of evidence, for example:

2026-10-14, shop/ledger-api, OOMKilled -> limits.memory 128Mi -> 160Mi -> verified

A lesson is history, not a fact about the cluster now: Nightshift is told to confirm it against this run’s reads. The alert’s panel shows which lessons the run opened.

An incident is remembered as one lesson when it closes: its root and the kind of failure there (taken from the evidence, not the model’s words), the symptoms upstream, the path between them, the fix on the root and whether it held, and how long it lasted. The same root, failure and fix again within 30 days is one more occurrence of the same lesson.

The next incident recalls past ones by their shape on the service map: the same root first, then the same kind of root failing the same way among the same services, then the same symptoms. Up to 5 are listed; opening one counts toward the same 3 a run may open.

On the incident page:

  • Seen before - for an admin, the similar past incidents and why each matched; for everyone else, only the ones the run opened.
  • What we learned - written by Seylo when the incident closes, from its facts: the root, the path, the fix and whether it held. Not the model’s words.

Nightshift → Memory (admin):

Control What it does
The list Newest first. Filter by Cluster, Namespace, Kind (alert or incident), Outcome, and From / To
Open The whole lesson
Delete One lesson, after a confirmation. Nightshift stops using it from the next investigation; investigations that already cited it keep their record
Delete all Every lesson, now and for good. Type delete all memory to confirm. Export first if you want a copy
Remember lessons On or off. Turning it off asks first: Nightshift stops storing lessons and stops using the ones it has. Nothing stored is deleted
Keep for (days) 30 to 730 days after a lesson last recurred; 365 by default. A shorter period deletes older lessons at once
Export Every lesson and the settings, as JSON (up to 5,000 lessons)

Nightshift → Settings → Memory shows whether memory is on, and links here. Every delete, export and change of setting is recorded in the audit log, without the lesson’s text. When your organisation leaves Seylo, its memory is deleted with the rest of its data.

When one fix has verified at least 3 times for the same alert and held more often than not, Nightshift → Memory offers Suggest a skill beside it. It drafts a SKILL.md - the goal, the workflow, what the findings mean and the fix that held, scoped to that alert and its namespaces - and opens it in the skills editor, unsaved. An incident shape that held 3 or more times drafts a skill scoped to the root service.

Nightshift never saves or adopts a skill itself. You read it, change it, and save it, or close it.