Ryker / Docs / Memory and learning
Memory and learning
Ryker can carry useful context from one conversation to the next: a team decision, a past investigation, or a correction you gave it. You can see what it kept, where it came from, and whether it should still be used.
On this page
Ask it to remember something
@Ryker remember for the team: keep the old billing API until the mobile rollout finishes. Keep this for 30 days.Ryker shows the proposed fact, where it applies, and when it expires. A configured operator confirms the card to save it. Facts can apply everywhere or be limited to a workspace, repository, or conversation. The card shows which. Ask “what do you remember here?” to see the facts available in that conversation.
If Ryker asks an operator a question and says it will remember the answer, that answer saves the fact without a second click.
Teach it through feedback
Tell Ryker what to do differently: “Always show the before-and-after replica counts in these reviews.” An operator’s explicit correction can be saved as guidance for later work in that conversation. The correction is already an instruction, so it needs no second confirmation.
Ryker shows what it saved and where it applies. Remove the guidance from that card, or manage it under Instructions. A reaction on a message alone does not save guidance.
Learn from conversations
Background learning reads retained messages in small batches and updates useful topics: a rollout decision, an intended configuration, or an unresolved problem. Each topic links to its source messages and keeps its earlier versions. A batch of ordinary chatter may produce no update.
Learning is on by default. It runs independently of replies, so Ryker can build knowledge while staying silent, including in an observe-only channel. Learning itself never posts a message.
Turn it off under Memory › Learning to pause processing. Messages wait, and existing knowledge stays. Under Memory › Learned, open a topic to inspect its sources, update history, and the request and answer behind each update.
Use memory in a new request
Ryker begins with a small selection of relevant facts, guidance, recent excerpts, and conversation summaries. It can search further by text, kind, source, or date, and follow an excerpt back to the original conversation.
The current request, safety rules, fresh evidence, current repository content, and configuration take priority over remembered context. A previous incident can suggest where to look, but Ryker needs new evidence to establish the cause or current health.
Recall respects access. Private-channel context stays within its permitted scope, and material from another channel requires current membership. Memory cannot grant permission to act.
What the memory pages contain
| Record | What it keeps |
|---|---|
| Source excerpts | The original words, source, and time of a message. These are available before background learning catches up. |
| Learned topics | The current understanding of a subject, its sources, corrections, open questions, and previous versions. |
| Conversation summaries | Where a conversation stands when work is accepted. Older summaries fold into weekly rollups without another model call. |
| Cases | Completed work: the problem, established cause, attempts, outcome, and sources. Related incidents remain separate cases. |
| Facts | Information people explicitly confirmed. Background learning cannot change it. |
| Findings | Investigation conclusions and the evidence behind them. |
| People | Context about teammates. |
Write team instructions
Use Instructions in the console for directions you want Ryker to follow. The installation-wide field applies everywhere. A channel can add instructions that take priority where the two disagree. Each field accepts up to 2,000 characters.
Ryker records the instructions sent with each model call. Saved changes apply when it prepares the next call; a call already prepared keeps its original instructions. The model cannot edit these instructions, and they cannot expand its permissions.
Set preferences and guidance
Preferences cover check depth (quick, standard, or deep), reply detail (concise, standard, or detailed), and where to reply: follow the conversation, prefer a thread, or prefer the channel.
Person settings take priority over channel settings, then repository settings, then workspace settings. Reply location supports person, channel, and workspace scope.
Guidance is advice in your own words. Personal guidance follows the person across channels; channel and workspace guidance hold team conventions. Manage both preferences and guidance under Instructions, where they can be paused, resumed, or deleted.
Correct or remove stale knowledge
Memory › Facts flags duplicate facts and facts that nobody updated, recalled, or reviewed in 30 days. Keep, edit, merge, or forget them there or in Slack App Home. The review never changes a confirmed fact on its own.
Forgetting erases the value; replaced and removed values are not copied into the audit record. If a learned topic loses one of its source messages, Ryker stops using it until you rebuild it from messages you choose.
How long memory lasts
Conversation memory lasts 90 days by default, set under Settings › Data retention. Reading a memory does not extend its life. Changes to the source message, deletion, or loss of access can make it unavailable sooner.
Confirmed facts follow the expiry date shown on their card, even when they apply everywhere. An answer Ryker explicitly remembers for the installation can have no expiry. Facts scoped to a workspace, repository, or conversation can also be removed after the conversation-memory retention period since their last change or review. Recalling a fact does not reset that period.
Cases have no routine expiry, but deleting their source message erases the text of every case built from it.
When a reply seems to ignore or misuse memory, inspect the request’s recorded context in the console to see what was actually included.