Your AI Doesn't Know Colour.
Now It Can.

Ask an AI for a colour palette and it can produce something beautiful in seconds. Ask it why those colours are right, and you find out how much of the answer was invented on the spot.

That is not really a colour-generation problem. It is a colour-knowledge problem.

Colour intelligence is the layer between a language model and a colour decision: provenance, measurement, cultural context, physical constraints, and honesty about what can and can't be claimed. Colour Memory is building that layer. It connects to AI assistants through MCP and gives them documented colour, colour science, accessibility data, and material information, instead of asking the model to invent the answer from its training data.

The AI reasons. Colour Memory verifies. The source remains the authority.

As of August 2026, Colour Memory holds more than 42,000 documented colours across 60+ research archives. It is not another palette generator, a paint catalogue, or a collection of AI-generated colour names.

The dangerous hallucinations aren't the ridiculous ones

A completely absurd colour answer is easy to reject. The dangerous one is a historically plausible hex, attached to a plausible date, described in fluent specialist language. It looks exactly like expertise. That's why provenance matters more than confidence.

Here's an actual receipt, reproducible

We ran this exact query against the live Colour Memory system on 12 August 2026: "Find a colour record grounded in an official Canadian government symbol."

Notice what's missing from that claim. It does not say #7B4397 is Ontario's official government colour, because no government source specifies that exact hex. It says amethyst's official status is documented, and gives a representative digital colour for the mineral. That's a deliberate distinction, not an oversight.

A source isn't the same thing as a measured hex.

A museum can prove an object existed. A register can prove a cultivar description. A government can prove the official status of a symbol. None of that automatically proves a modern six-digit hex is the uniquely correct digital representation of it. Colour Memory records those as two separate questions and is designed to keep them separate in its responses, rather than quietly collapsing "documented" and "precisely measured" into one confident-sounding answer.

Ask it something a general model was never trained to answer

"Find a documented sweet pea cultivar colour from an official register."

Not a vague gesture at "sweet pea pink." A specific cultivar, tied to a named register.

Sometimes the best answer is no

Search more loosely for "sweet pea pink" and Colour Memory also surfaces an entry with that exact name, evidence grade E: poetic or literary inference only. Its own record states plainly: do not cite a primary source, none is recorded for this entry.

Same flower family. One entry backed by a real institutional register. One flagged, in its own metadata, as having no source at all. Colour Memory doesn't quietly smooth the weak one up to sound as credible as the strong one.

A system you can trust has to be able to disappoint you.

If every query magically returns an authoritative-sounding match, the system is probably laundering uncertainty into confidence rather than reporting it honestly.

The model doesn't need to know everything

No language model should be expected to reliably reproduce every specialist archive, registry, and regulatory standard from its weights alone, any more than a doctor should be expected to recall every drug interaction from memory instead of checking. The better architecture separates the jobs:

The language model reasons, converses, and synthesises.
Colour Memory supplies the evidence, the colour science, and the claim boundaries.
Primary sources remain the authority. Colour Memory cites them, it doesn't replace them.

That division of labour is bigger than colour. It's a template for how expert AI should work in any domain where getting it wrong has consequences.

What's actually in the archive

The 60+ archives cross disciplines because colour does: historic material culture, flowers and horticulture, fashion and textiles, brands and trademarks, government and regulatory records, architecture and interiors, and specialist technical data. Alongside the historical and cultural layer, Colour Memory now also carries substrate compatibility data (which coatings actually work on which materials), Light Reflectance Value, and thermal performance data, where available, the gap between a nice colour and a colour that's actually specifiable for a real build.

What can Colour Memory do that a general model can't?

Capability	General-purpose AI	Colour Memory
Generate colour ideas	Yes	Yes, via a connected AI
Retrieve a graded evidence record for a colour	Not inherently	Yes
Attach an explicit claim boundary (what it must not say)	Not inherently	Yes
Separate "source confirmed" from "hex measured"	Not inherently	Yes
Calculate contrast/accessibility rather than estimate it	Tool-dependent	Yes
Persist a colour project across sessions	Platform-dependent	Yes

"Not inherently" is doing real work in that table. A general model can sometimes recall a genuine source or express uncertainty. What it isn't structurally bound to is an evidence record that defines, in advance, what's permitted to be claimed and what isn't.

It began with a simple frustration

Colour Memory was built by a working photographer and filmmaker who kept running into the same problem: AI could describe colour beautifully and confidently, but couldn't reliably say where any of it came from. Once that started to matter, the deeper issue became obvious. Colour knowledge is scattered across museum catalogues, horticultural registers, government records, and scientific datasets that never talk to each other, and no AI model is built to reconcile them. Colour Memory exists to be the layer that does.

Is this a Pantone replacement?

No, and the comparison is useful for saying what Colour Memory actually is. Pantone is a colour communication standard: it helps two people specify the same colour. Colour Memory is a colour-intelligence layer: evidence, colour science and constraints for AI. It helps an AI answer where a colour came from, whether a historical claim holds up, and whether the confidence attached to it is actually earned.

Can't ChatGPT already do this?

Sometimes, yes, it will land on exactly the same colour. The problem is you don't know why it got there. Did it retrieve a real fact? Remember one approximately? Infer something plausible? Invent a citation-shaped explanation after picking a colour that just looked right? Colour Memory doesn't claim a general AI can never arrive at the same answer. It makes the route to that answer inspectable, which a fluent guess never is.

AI doesn't need another billion invented colour combinations. It can already generate those on its own. What it needs is a way to distinguish the documented from the plausible, the measured from the inferred, and the known from the merely convincing.

That's Colour Memory. Colour with receipts.