Understand which context you are changing
Purpose and Working instructions are included in the agent’s working context. An instruction in chat can guide the current conversation, but it becomes persistent agent memory only after a successful explicit save.
| Context | What belongs here |
|---|---|
| Agent profile | Purpose, Working instructions, competencies, responsibilities and expected work products. |
| Rules & memory | Your reusable guidance, preferences, methods, contextual facts and examples for this specific agent. Any subject is supported; there is no required list of topics. |
| Project operating rules | How this agent communicates and behaves in the selected project. |
| Company and project knowledge | Authorized source-linked operating knowledge and evidence about the work. It is managed separately from an individual agent’s rules. |
| Current task and map step | The immediate requested result, supplied evidence, fixed inputs and action boundaries. |
Add or edit guidance in Settings
Expand the agent’s Purpose, Working instructions and competencies. These describe its responsibility; saved personal rules are listed separately in the same window.
- 1Open the agent’s settings
Choose Rules & memory. The same memory can be opened from the Agent rules and memory button in the agent chat.
- 2Inspect the existing profile and rules
Expand Profile, instructions & competencies. Review the active and paused entries before adding overlapping guidance.
- 3Write a named rule
In Teach your agent, enter Rule name and Rule content. Explain the requirement and when it applies; add examples if useful. A rule can contain up to 1,800 characters, with a name of up to 100 characters.
- 4Review the result and save
Use Remember for a new rule or Save rule when editing. If related or conflicting guidance is found, inspect the existing text and the exact proposed result. Choose an option before the saved memory changes.
Manage the same memory through chat
Ask “Remember: for incident reviews, show the impact, evidence timeline and recovery options.” You can also ask to change a named rule or forget one. These requests may concern any competency, work method, content style or useful context.
The agent presents a proposed change. Use Remember, Forget rule or the offered review option to apply it. A chat reply saying that a rule should be remembered is not confirmation of storage. The interface reports Saved only after the server confirms the change.
Chat and Settings share the same saved entries. Return to Rules & memory to inspect, edit, pause or remove the result.
Handle overlap and contradictions explicitly
A new or changed rule is reviewed against this agent’s saved guidance. Exact duplicates do not create another copy. Semantic review can identify similar wording and conflicting requirements; it can make mistakes, so read the proposal.
Combine with shows the complete proposed merged text. Replace shows the replacement. Keep as a separate rule retains distinct guidance when that is your intention. Nothing is merged, replaced or removed merely because a review was displayed.
Different conditions can coexist: concise chat replies and detailed analytical reports are not inherently contradictory. Make the conditions explicit. If rules change elsewhere while you are reviewing, reload the current memory and review again.
How memory is used without growing every request
Only active entries are used. Small memories are passed completely. For larger memories, conditional rules that have a retrieval index can be selected by the current task or workflow step. General rules, older unindexed guidance and uncertain matches are retained.
The applicability explanation shown on a reviewed conditional rule describes when it is relevant. Existing entries remain available; they are not silently removed or assigned a restrictive topic.
Specific current task requirements and project instructions take priority over general defaults. A concise response preference does not automatically shorten a requested detailed report. Memory cannot grant permissions, select extra recipients, authorize sending or replace approval.
Saving a new or changed rule can use AI for review and indexing. Applying an already reviewed choice does not repeat that analysis. Small memories do not require a separate retrieval request before each action; larger-memory retrieval reuses bounded cached results.
Pause, remove and revise as you learn
- Pause a rule to retain it while excluding it from future work; activate it again when appropriate.
- Edit a rule when experience reveals a clearer requirement. Remove it when it should no longer guide this agent.
- Saved memory is limited to 32 entries and 12,000 characters across their names and content. Prefer compact, specific guidance and combine genuine overlap after review.
- This is controlled working memory, not automatic retraining of model weights. Source documents and tool responses do not silently write new owner rules.
- Compare actual outputs before and after a rule change. Evaluate evidence, requirement coverage, errors, time and usage rather than assuming that more saved rules always improve results.