Halon's memory
Halon keeps short facts about your setup, such as which database is production or which service handles payments, so the next chat or investigation starts from what your team already knows. Every memory is a hunch Halon checks against live results, never proof, and it says so when it relies on one nobody confirmed.
Where memories come from
Section titled “Where memories come from”| Source | When | Starts as |
|---|---|---|
| An incident | When the incident ends, Halon reads what was found and said and keeps at most three lessons | Unconfirmed |
| A late correction | An answer marked wrong after its incident ended teaches at most one lesson about what Halon got wrong | Unconfirmed |
| A postmortem | When a postmortem is marked completed, Halon reads it against what the incident taught and keeps anything new | Unconfirmed |
| A chat or an investigation | Halon works out a lasting fact while answering | Unconfirmed |
| You, in a chat | You ask Halon to remember something | Confirmed by you |
| You, on the Memory page | You click Add memory | Confirmed by you |
Halon never saves the same fact twice, even in other words, and never learns again a fact someone marked not right. It also refuses two kinds of fact on its own: a reading that will have changed by the next incident (an error rate right now, a timestamp, a count of what is failing) and anything about a person (a name, a handle or an email address). Who owns what belongs in the catalog.
An investigation saves and disputes memories as Firefight’s own Halon agent, which holds Memory create and update by default. Each save and dispute is in Activity under its name. To stop investigations changing memory, revoke those grants under Gateway → Permissions.
When something says otherwise
Section titled “When something says otherwise”When Halon learns a fact about something it already remembers a fact about, it checks whether the two say the same thing, contradict each other, or are unrelated. The same fact in other words is not saved again. When they contradict, Halon keeps the new fact, stops using the old memory (it becomes Disputed) and asks which is right straight away, where it found the contradiction:
- In a chat on the dashboard, a Which is right? card appears under the answer. It quotes the old memory, says where it came from and who confirmed it, and says what showed otherwise.
- In Slack, the incident’s channel gets a note in the thread of the chat or investigation, as for any dispute.
- On the Memory page, the old memory waits under Disputed with what showed otherwise.
A card reads like this.
I used to remember “web deploys from main” (from your chat on Oct 3, 2026, confirmed by Ana). The deploy log of web shows “web deploys from the release branch”. Which is right?
The same card appears in a dashboard chat when Halon reads a live result that contradicts a memory and disputes it.
| Answer | What happens |
|---|---|
| Still right | The old memory is confirmed by you, and Halon stops using the new fact |
| Not right | The old memory is marked not right, and Halon uses the new fact from then on, confirmed by you |
| Correct | Opens a form for what is right instead. Your correction replaces both, confirmed by you |
The card then says how it was settled. You can answer from the card, from Slack or from the Memory page (where Confirm keeps it as still right), and the answer settles both facts wherever you give it.
Deciding in Slack
Section titled “Deciding in Slack”Halon asks the incident’s channel about what it learned. You see one message when an incident ends, one when its postmortem teaches something new (titled after the postmortem), and a short note in the thread of a chat or investigation when Halon learns a fact or disputes one while working on the incident.
Each memory in the message carries the actions its state allows.
| Action | What happens |
|---|---|
| Confirm | The memory is confirmed by you, and Halon reads it that way from then on |
| Still right | Shown instead of Confirm on a memory that is disputed or possibly outdated. It confirms it as still right |
| Not right | Halon stops using the memory and never learns that wording again |
| Correct | Opens a form for what is right instead. The old wording is kept as not right, and your correction replaces it, confirmed by you |
Once someone decides, the message says who, and the buttons go. A message only ever shows the memories it was posted with, and one deleted on the Memory page drops out of it. If the channel has been archived, nothing is posted and the memories wait on the Memory page.
Weekly reminders
Section titled “Weekly reminders”Memories nobody confirmed are easy to forget, so Halon reminds the people who can answer, once a week on Mondays.
| Who taught it | Who is reminded |
|---|---|
| You, in a chat | You, by direct message, about your own memories only |
| An incident, its postmortem or an investigation | The incident’s channel |
| Halon, with no incident to ask in, or one whose channel was archived | Each workspace admin, by direct message |
A reminder only goes out when something has waited at least three days, holds at most five memories, and never shows a memory the last reminder to the same person or channel showed. Each person and each channel gets at most one a week. It opens with how many memories wait and why it matters.
Halon picked up 2 things from your chats that nobody has confirmed yet. Confirming keeps them in use, and anything wrong stops Halon repeating it.
Each memory carries Confirm, Not right and Correct, and the message links to the Memory page, where everything waiting is listed. People who have no Slack account in the workspace get no reminders, and their memories wait on the Memory page.
When a postmortem disagrees
Section titled “When a postmortem disagrees”A completed postmortem confirms the lessons it agrees with, credited to whoever marked it completed. When it was completed through the API or by an agent, the memory reads as confirmed by a postmortem rather than by a person.
A postmortem never overrules a person. When it contradicts a lesson nobody confirmed, Halon marks that lesson not right and, when the postmortem says what is right instead, remembers that. When it contradicts a memory a person confirmed, the memory becomes Disputed with what the postmortem says instead, and a person decides.
The Memory page
Section titled “The Memory page”Open Memory in the dashboard’s sidebar, under Halon. The page has two tabs, Memories and Instructions. The Memories tab shows how many memories are waiting on a person. Each memory shows what it is about, where it came from, who stands behind it and how often Halon used it. Being handed a memory at the start of a chat or an investigation counts as one use for that chat or investigation. Search memories narrows the list by its words.
| Filter | What it holds |
|---|---|
| In use | Everything Halon reads |
| Unconfirmed | What Halon learned that nobody has confirmed yet |
| Possibly outdated | Memories about something renamed, archived in the catalog, or gone from the map. Halon still uses them, flagged. When a resource comes back on the map, the flag goes |
| Disputed | Memories a live result, a postmortem or a newer fact contradicted. Halon stops using them until someone decides |
| Expired | Unconfirmed memories nobody confirmed in time, when your workspace sets a window |
| Rejected | Memories someone marked not right, kept so Halon never learns them again |
Confirm, Correct and Not right work as they do in Slack, and the list stays on the filter you were working through. To delete a memory, open the row’s menu and choose Delete. Firefight asks first, and deleting is for good. Deleting a rejected memory lets Halon learn that wording again, so mark it not right instead when you only want it gone from use.
Every member can add, decide on and delete memories. An agent or a service key needs the Memory permission, see Permissions.
Confirmed memories also turn up in the dashboard’s search, Search the map, catalog and memory, and in the search_map MCP tool. Choosing one opens it on the Memory page.
A memory about a resource you cannot see on the map is missing for you everywhere, on the Memory page, in search and in Slack, and Halon leaves it out when it works for you.
Instructions for Halon
Section titled “Instructions for Halon”Memories are facts. Instructions are how you want Halon to work, such as which logs to check first, and they live on the Instructions tab of the same page. See Instructions for Halon.
Letting unconfirmed memories expire
Section titled “Letting unconfirmed memories expire”By default Halon uses an unconfirmed memory until someone decides on it. An admin can change that under Settings → Workspace, in the Halon card, with Stop using unconfirmed memories.
| Choice | Effect |
|---|---|
| Never, use them until someone decides | The default |
| After 30, 60, 90 or 180 days without a confirmation | A memory nobody confirms in that time moves to Expired and Halon stops using it |
An expired memory can still be confirmed on the Memory page, which brings it back into use. One Halon learns again starts waiting afresh. Agents can change the same setting with update_workspace_settings over MCP, as memory_expiry_days.
When a memory changes during an investigation
Section titled “When a memory changes during an investigation”An investigation starts from the memories that applied when it began. If someone disputes, corrects, marks not right or deletes one of them while it runs, Halon is told at its next step and stops relying on it.