Skip to main content
Memory is the part of Atlas, Rootly’s AI layer, that holds what AI SRE has learned about your systems. Each learning is a context note: one short, reusable claim about a service, dependency, signal, or failure pattern. A later investigation can recall the note and test it against current evidence, so AI SRE reuses what the selected team already knows without treating an old conclusion as proof. AI SRE is the only Rootly AI feature that reads Memory, so you review notes from the Atlas page: open AI SRE → Atlas and select the Memory card. If your sidebar still shows AI & Agents, the card is under AI & Agents → AI SRE; see Where AI Settings Live.
Memory may not be enabled for every team. The Rootly account team turns it on, along with the features that propose notes after investigations, recall them, and activate them automatically. The Memory card appears on the Atlas page only when AI SRE is enabled for the selected team; see Getting Started. Viewing Memory needs an Incident Response seat. Editing notes or changing their status needs the Incident Response Owner or Admin role with that seat. See Manage User Permissions for the full matrix.

​
How Memory Works

Notes reach Memory in two ways:
  • After an investigation. AI SRE reviews findings from runs listed as Root cause identified or Best effort on the Investigations page alongside your existing notes. Best effort maps to Contributing factor identified on the tab for teams using the six-stage flow. On the default flow, a Best effort run can show Inconclusive — needs human on the tab and still propose notes. AI SRE can propose notes for what the investigation established, and it can retire or replace existing notes (see What AI SRE Changes on Its Own). Runs listed as Inconclusive, Blocked or Could not investigate on the Investigations page propose nothing, and neither do failed or cancelled runs.
  • From past incidents. Rootly can seed Memory with proposed notes drawn from recent incidents that have a published retrospective and allow Slack history collection. It reads each incident’s Slack channel and retrospective, keeps only claims it can support with verbatim quotes, and drops any claim the retrospective contradicts. Ask your Rootly account team whether this has been run for the selected team.
Every note starts as Proposed, and AI SRE recalls only Active notes. A useful note is:
  • Specific to a durable service, topology, mechanism, signal, or failure pattern.
  • Likely to help with a future alert.
  • Grounded in what the source investigation or incident established.
  • Written as a hypothesis the next investigation can verify.
One-off timestamps, customer or user identifiers, incident recaps, generic SRE advice, and instructions to acknowledge, ignore, restart, or roll back a system don’t belong in Memory. Rootly generalizes findings into reusable system behavior and discards a candidate that only makes sense for the original customer, person, or incident. An investigation that produces no note is a normal result.

​
Private Incidents

Private incidents never feed Memory. An investigation of a private incident, or of an alert linked to any private incident, doesn’t propose, grade, or automatically activate notes. If an incident later becomes private, notes sourced from it, and from alerts linked to it, drop out of the Memory list, the card count, version history, and direct note links, and AI SRE stops recalling them. This applies even to people who can view the private incident, because Memory is shared across the selected team.

​
How a Note Becomes Active

AI SRE never sees a Proposed note. A note becomes Active in one of two ways. A person activates it. An Owner or Admin reviews the note and selects Activate, or Reactivate on a Quarantined note. The note is Active immediately. It qualifies for automatic activation. Where Rootly has turned on automatic activation for the selected team, a note can become Active without anyone approving it. The rule is a fixed count:
  1. A later investigation searches Memory and comes close to the Proposed note. AI SRE isn’t shown the note.
  2. After the investigation finishes, Rootly grades whether the note would have helped reach that investigation’s finding, on a scale of 0 to 100.
  3. Once at least two separate investigations have each graded the note 70 or higher, it becomes Active. Several grades within one investigation count once.
Only Diagnostic Hint, Known Failure Mode, Topology Fact, and Historical Context notes can activate automatically. False Positive Pattern notes, notes with any other category, and notes whose Type someone changed while editing always wait for a person. A proposal that never comes close to a later investigation stays Proposed until someone reviews it. Automatic activation is a Rootly-managed setting, and the Memory page has no control for it. Contact your Rootly account team to confirm whether it’s on for the selected team, or to turn it on or off.

​
What AI SRE Can Remember

Each note has one category. The Memory list shows it on every row, or General when a note has no category. A false-positive note should explain how to verify the pattern from evidence. It shouldn’t tell AI SRE to dismiss or acknowledge an alert without checking.

​
Review Context Notes

1

Open Memory

Open AI SRE → Atlas → Memory (AI & Agents → AI SRE → Memory if your sidebar doesn’t have an AI SRE item).
2

Find a Note

Search by title, summary, or category, or narrow the list with the All categories, All sources, and Any status filters. The source filter separates notes Rootly created (AI) from versions a person edited (Manual).
3

Review the Note

Select a note to open it. Read its Knowledge, the summary AI SRE sees when it recalls the note. Notes seeded from past incidents also show Scope, such as the service, environment, or alert the note applies to, and Evidence, with quoted excerpts that link to the incident’s Slack channel or retrospective.
4

Choose a Status

Activate a reliable note, deprecate an Active note that no longer helps, or quarantine a note that’s wrong or unsafe to recall.
5

Correct the Wording

Select Edit to change the Title, Type, or Knowledge, then select Save new version. See Editing and Version History.
The actions on a note depend on its status: Everyone who can view Memory sees History. Only Owners and Admins with an Incident Response seat see the other actions. Memory has no delete, no reject, and no way to write a note from scratch. To turn down a proposal, quarantine it or leave it Proposed. To give AI SRE guidance it should follow in every investigation, write Instructions.
Memory is shared across the selected team. Everyone who can view the Memory page sees every note from an eligible source, whatever its status, and can open its earlier versions in Version history. The page doesn’t re-filter a note’s title, summary, scope, or evidence by whether the viewer could run the AI connector or private agent capability the finding came from. Treat Memory viewers and future investigations as the audience for anything learned from a public incident or alert.

​
Memory Statuses

​
What AI SRE Changes on Its Own

Where Rootly has turned these on, AI SRE updates Memory after investigations without anyone acting, so you may see statuses you didn’t set:
  • Deprecate, quarantine, or replace. When AI SRE reviews a finished investigation, it can deprecate a note that no longer helps, quarantine one the investigation showed to be wrong, or replace a note with a reworded version.
  • Quarantine after recall. After an investigation, AI SRE reviews each note it recalled. It quarantines a note when live evidence contradicted it or when the review finds the note did more harm than good.
When AI SRE replaces a note, the original becomes Superseded immediately and the replacement starts as Proposed. If the original was Active, AI SRE stops recalling that learning until someone activates the replacement or it qualifies for automatic activation. Filter by Proposed from time to time to catch replacements of notes you rely on. The Memory page doesn’t show who changed a note’s status, so a Quarantined badge can come from a person or from AI SRE.

​
Editing and Version History

Notes are versioned so a correction doesn’t rewrite the record. Edit opens Edit context note with three fields: Title, Type (the category), and Knowledge. Saving never changes the note in place. It creates a version in the same lineage and marks the previous version Superseded. AI SRE reads a recalled note’s title, category, and Knowledge, and those fields also shape how the note is matched. An edit changes what AI SRE sees. The saved version:
  • Is credited to you. Its source shows your name instead of AI, and it appears under the Manual source filter.
  • Stays Active only if the previous version was Active. Editing a Proposed, Deprecated, or Quarantined note produces a Proposed version, so editing a Quarantined note puts it back up for review.
  • Loses automatic activation if you change Type. Clearing Type removes the category, and the note shows as General.
  • Keeps the previous version’s scope and evidence.
  • Moves to the top of the list, which shows the newest notes first.
After you edit an Active note, AI SRE can recall the edited version once Rootly finishes indexing it. Select History to open Version history. The button shows how many versions the note has, and the history lists each one with its status, source, and summary. It lists versions only: activating, deprecating, or quarantining a note updates the current version without adding an entry. If a note’s claim is wrong rather than badly worded, quarantine it instead of rewording it.

​
How AI SRE Recalls Notes

  • AI SRE recalls only Active notes whose source isn’t private. Proposed, Deprecated, Quarantined, and Superseded notes never reach it.
  • At the start of an investigation, and while it works, AI SRE can look up the Active notes closest in meaning to the alert or incident. Rootly ranks them by that closeness and by each note’s track record in earlier investigations, then passes AI SRE the top few.
  • Recall has no minimum match. If the selected team has Active notes, AI SRE receives the closest few even when none of them is closely related.
  • AI SRE treats every recalled note as a hypothesis to check against live evidence. It can’t cite a note as evidence, and it’s told never to conclude a root cause or declare a false positive from a note alone.

​
Recent Alert Context

Alert investigations also read recent AI SRE conclusions on related alerts. That lookup is separate from Memory and can’t be curated here; see Past Alerts, Incidents and Conclusions. For everything else an investigation reads, see Evidence Sources.

​
Keep Memory Reliable

  • Activate a note only when its claim is specific, reusable, and supported.
  • Quarantine proposals that hold on to customer, user, incident, or request identifiers instead of a durable system behavior.
  • Prefer a verifiable signal over a prescribed response.
  • Deprecate an Active note that’s accurate history but no longer helps. Recall has no minimum match, so an outdated Active note can still reach AI SRE.
  • Quarantine a note that contradicts current evidence or could mislead a responder.
  • Recheck notes after service renames, topology changes, or observability migrations.
  • Use Instructions for practices every investigation should follow, and Memory for context recalled only when it matches.

​
Data Handling

Context notes and their version history are stored as durable Rootly records. The 90-day expiry for AI session messages doesn’t apply to them. Deprecating or quarantining a note removes it from recall but doesn’t delete or hide the note or its earlier versions, and the Memory page has no delete action. Notes belong to the selected team, and investigations on other teams cannot recall them. The evidence used to create or recall a note can also appear in AI conversation history, model traces, and connector-specific records. See Data Privacy for Rootly AI for the complete storage and retention boundaries.

​
Troubleshooting

Memory may not be enabled for the selected team, and the card appears only when AI SRE is enabled. Confirm that you’ve selected the right Rootly team and hold an Incident Response seat on it. If both are correct, contact your Rootly account team.
AI SRE recalls only Active notes. Review the proposal and activate it if the claim is reliable and reusable. Where automatic activation is on, an eligible proposal can instead become Active after two separate investigations each grade it 70 or higher; proposals that don’t meet that rule stay Proposed.
AI SRE can deprecate, quarantine, or replace notes after an investigation, and it quarantines recalled notes that live evidence contradicted. If it replaced an Active note, the replacement waits as Proposed; activate it to restore recall.
Deprecate appears only on Active notes. To withdraw a Proposed note, quarantine it. A Quarantined note offers only Reactivate, so leave it Quarantined if it should stay out of use.
Memory keeps Deprecated and Quarantined notes for review, and Superseded versions in Version history. Filter by Active to see only the notes AI SRE can recall.
Review the note’s knowledge and evidence, then quarantine it if the claim is misleading. AI SRE should recheck recalled context against current evidence; if the result leaned on the note anyway, rate it under How accurate was this investigation? See Reading the Result.

Atlas Overview

Rootly’s AI layer and the ways people and applications use it.

Knowledge Graph

The services, infrastructure, and relationships AI SRE uses during investigations.

Instructions

Guidance AI SRE follows in every investigation, unlike notes it recalls only when they match.

Evidence Sources

Everything an investigation reads, and the access boundaries around it.

Reading the Result

Read a finished investigation and rate its accuracy.

Data Privacy for Rootly AI

How Rootly stores, retains, and shares AI data, including Memory.

​
Frequently Asked Questions

No. AI SRE proposes a context note only when the investigation established a durable, reusable learning about your systems. No note is a valid outcome when the finding is one-off, customer-specific, procedural, or too uncertain.
No. A note supplies a hypothesis or a shortcut for where to look. AI SRE must verify it against the current alert’s or incident’s live evidence before using it in a diagnosis, and it can’t cite the note as evidence.
No. You can edit an existing note, which saves a version credited to you, but you can’t create one from scratch. For guidance AI SRE should always apply, write Instructions, or add instructions to an investigation rule.
It depends on automatic activation. When it’s off for the selected team, every note waits for a person. When it’s on, eligible categories can become Active after two separate investigations each grade the note 70 or higher, while False Positive Pattern notes, notes with other categories, and notes whose Type was edited still wait for a person.
No. Private incidents, and alerts linked to any private incident, don’t propose, grade, or automatically activate notes. If a source incident later becomes private, notes sourced from it drop out of the Memory list, version history, direct note links, and recall, even for people who can view that incident.
No. Memory uses statuses and version history instead of deletion. Deprecate an Active note that’s obsolete, or quarantine one that’s misleading or unsafe to recall.
No. Only AI SRE recalls context notes. The Rootly Agent can use other parts of Atlas, such as AI connectors and private agents.