Reflection

Airic does not silently update hidden memory. After work is done, agents reflect on the process and propose changes as Markdown diffs you review and approve.

Reflection questions

After each run, Airic asks:

• What went wrong?
• Which context caused the mistake?
• Was the process unclear?
• Was the document type missing a requirement?
• Should this human judgment become a precedent?
• Should the pack be improved?

Patches, not prompts

Improvements become spec patches — visible diffs against methodology documents:

@@ -12,6 +12,11 @@
   When framing a customer problem,
   always restate it as a Job-To-Be-Done
   before generating solution candidates.

+ ## Precedent (added 2026-06-27)
+ If the user describes a workflow,
+ extract the trigger, the desired
+ outcome, and the current workaround
+ before proposing any solution.
- // TODO: revisit framing rules

You approve or request changes before the pack is updated. Future runs inherit the improved methodology.

Human-approved learning

Airic learns by editing methodology documents — with human approval.

This is the opposite of silent memory updates:

Approach	Airic
Hidden prompt edits	Visible Markdown diffs
Automatic behavior change	Explicit approval step
Opaque learning	Auditable precedent library

See the reflection section on the landing page for a live mock session, or compare approaches in vs Traditional agents.