A substrate
you can walk.
HeatherDB ingests raw input and self-organises it into an associative substrate. No schema. No training. No annotation. What you get is a navigable geometry of co-occurrences — the thing every other AI architecture either hides, fakes, or forces a human to build by hand.
LLMs hide the substrate in opaque weights. Vector DBs store points without edges. Knowledge graphs require you to author the graph manually before you can query it.
HeatherDB lets the substrate form itself, then exposes it directly.
Build it once. Read it three different ways.
See what's in your substrate
Click any node in the substrate. See its strongest associations. Click any of those, walk outward. You're traversing the geometry the way human memory traverses itself — by association, not by query. The first interpretable AI substrate. No backprop forensics required.
- Onboard a new hire by letting them wander the team's accumulated substrate
- Audit a brand archive by walking what's in it before you trust it
- Discover what a corpus actually 'knows' without writing a single query
Score new content against the substrate
Paste any text. Heather builds its pair vectors and asks the substrate to recognise them. High score: this content sits inside the substrate's geometry. Low score: it doesn't. A forgery-resistant signature for any voice, brand, or institutional pattern — because the substrate's shape can't be imitated without the inputs that built it.
- Catch off-brand content before it ships
- Detect AI imitations that survive surface-style classifiers
- Behavioural baselines for security or compliance per-account
Ground any LLM in your specific substrate
Replace the retrieval layer in any RAG pipeline. Instead of returning K nearest documents, return the substrate's associations around the topic — what travels with what in your specific context. The LLM stops sounding like the internet average and starts sounding like you.
- Drop-in replacement for the vector-DB step in any RAG stack
- Personal-AI memory that compounds with use
- Multi-tenant LLM grounding where each tenant has substrate isolation
The empty slot in the data stack. Until now.
| Layer | Query primitive | What it gives you | What it can't |
|---|---|---|---|
| SQL | Relational query | Tabular facts | Anything unstructured |
| Vector DB | Nearest-neighbour | Point similarity | Show you the shape of the data |
| Knowledge graph | Authored traversal | Schema-bound facts | Form a graph from raw input |
| LLM | Generation | Plausible text | Show its working |
| HeatherDB | Walk / fingerprint / associate | Learned associative geometry | Generate text or store typed facts |
HeatherDB is not a replacement for any of these. It's the layer none of them occupies: the associative geometry that forms automatically from raw input and stays directly inspectable. Sit it next to your vector DB; pair it with your LLM; complement your knowledge graph. The substrate is the missing primitive.
The interpretability problem, solved by construction.
Mechanistic interpretability is the active frontier in AI safety research — and it's hard precisely because the substrate (model weights) wasn't designed to be readable. Reverse-engineering features from a 70-billion-parameter transformer is the work of thousands of researcher-hours per finding.
HeatherDB's substrate never went through that abstraction. Every co-occurrence is still locatable. Every basin is still queryable. Every edge traces back to specific writes. The substrate stays in its raw form.
For high-stakes deployments — financial services, healthcare, legal, regulated industries, EU AI Act-bound systems — this isn't a nice-to-have. It's the only architecture where audit and inspection are primitives, not bolt-ons.
The associations Heather surfaces map directly to the co-occurrences in your source material. No hidden reasoning step, no hallucinated bridge.
Remove a passage from the source, the corresponding associations weaken. Add a passage, new edges form. No retraining required.
Substrate isolation by tenant. Audit trails by construction. Forgetting-by-design is a database operation, not a fine-tuning campaign.
Walk a substrate. See what it's like.
The lexicon lab is one substrate built from Grimm's Fairy Tales. The same primitive works on any corpus — your team's writing, your customers' transcripts, your codebase, your archive. Start by walking ours.