Architecture philosophy
The principles. Not the algorithms. Enough to understand what kind of system Finch is — not enough to clone it.
Persistent learning
Every interaction adds to a substrate that survives the next session. Finch isn't a stateless prompt-completion machine — it's a state-accumulating one. The interesting work is in what state to keep, what to merge, and what to discard.
Cumulative capability
Yesterday's mastered skills are today's prerequisites. New topics build on top of validated foundations. The capability curve is a slow, longitudinal climb — not a per-prompt sample.
Structured memory
Memory isn't a single bag of vectors. It's layered: raw Q&A pairs, extracted skill patterns, an ontology-normalized knowledge graph, and a failures registry that captures what went wrong. Each layer answers a different kind of question.
Failure-driven refinement
When Finch fails, the failure is recorded with its category and a fix direction. Future attempts on similar tasks see the lesson before they start. Failure is a first-class signal, not a result to silently retry past.
Transfer learning
Concepts mastered in one domain seed adjacent domains. The system actively looks for patterns that should generalize — and tracks whether they actually do. Transfer isn't assumed; it's measured.
Ontology stabilization
A growing graph fragments without active maintenance. 'rest api' and 'REST API' and 'RESTful API' have to become one node. The system runs canonical merges nightly to keep concept density climbing instead of leaking into synonym duplicates.
Compositional reasoning
The substrate isn't just retrieval. Stored skills compose into new procedures. Stored facts compose into new conclusions. The point of the layered memory is that downstream code can ask 'has this been demonstrated before?' and get a structural answer, not just a similarity score.
SCOPE // What this list is not: an algorithm description, an architecture diagram, or an invitation to reproduce. The point of saying these aloud is to set the project's direction publicly — the actual mechanics live in the codebase.