Adaptive memory

What the copilot has learned about Fidelis Global

Every correction you make can become a learning — a small, reversible rule. Each rule is scoped (just this item, a surface-type, or brand-wide), injected into future work so one fix propagates to all similar output, and — once it proves stable through repeated reinforcement — offered to graduate into a structured record in the profile or style guide.

This item applies to one post or section only
A surface-type every image, caption, or carousel of that kind
Brand-wide injected into all brand knowledge & content
All Factual Preference Directive Style 9 active
Fidelis leads with enterprise relocation; individual moving is secondary; we operate in 30 countries.
Directive Brand-wide ×2 from: "We pivoted…"
Active
Household-goods insurance cap is ₹7.5L, not ₹5L.
Factual Brand-wide ×1 from: "Fix the coverage figure"
Active
Product shots: flat minimal illustration, no photoreal.
Style Surface: image ×3 from: "Make it an illustration"
Active
Avoid sports metaphors in captions.
Preference Surface: captions ×2 from: "Drop the 'home run' line"
Active
Carousel slide 1 must open with a bold stat hook.
Preference Surface: carousel ×4 from: "Lead with the 98% number"
Active
Primary CTA wording: "Get a move plan".
Directive Brand-wide ×5 from: "Standardise the button copy"
Active
Brand voice = confident, plain, no corporate jargon.
Preference Brand-wide ×6 → baked into Voice section v5
Graduated
Tagline: "Relocation made personal, in 15+ countries".
Directive Brand-wide replaced 2026-07-06
Superseded
This rule is stable — bake it into the style guide?

You've reinforced "Product shots: flat minimal illustration" 3 times across image generations. Rules this stable are better held as a first-class style-kit look variant than re-injected every time.

Rule
Product shots: flat minimal illustration, no photoreal
Style-kit look variant
image · flat-minimal
Graduating means it stops being separately injected — the variant carries it — and it stays an auditable pointer back to the corrections that formed it.
Learnings are additive and reversible — they never silently overwrite the profile or style kit. They layer in at generation time (that's how one correction propagates to all similar future work) and only graduate into canonical records on your approval.