OCRA — Ontology-Centric Reasoning Architecture

OCRA is a prototype for turning long, messy chat threads into structured OWL ontology artifacts. The current repo demonstrates the "create" pathway: ingest an artifact, generate modelling perspectives, build TBox/ABox ontology fragments, and merge them into a unified OWL output.

The motivating question is whether agent memory can be grounded in something more inspectable than a transcript or vector store: typed concepts, relations, provenance, and mergeable symbolic structure.

Current prototype status

This is a research prototype, not a polished product.

What works in the saved run:

• A Buddhist / Zen / phenomenology chat thread is stored in src/inputs.ts.
• The workflow decomposes the thread into ontology fragments.
• Fragment outputs are stored by the local viewer under src/viewer/.output/ when the viewer is run.
• A merged OWL artifact is committed as unified.owl.
• The committed unified.owl contains:
  • 134 OWL classes
  • 117 object properties
  • 147 named individuals
  • 782 rough direct assertion triples

What is still prototype-grade:

• The ontology is useful as an artifact, but not yet a validated production memory layer.
• Merge/repair behaviour still needs stronger consistency checks, provenance UX, and reasoner-backed validation.
• The repo contains historical debug scripts and experiment scaffolding that should be consolidated before broader public release.

Important files

• src/inputs.ts — source chat thread for the saved Buddhist / Zen run.
• unified.owl — merged OWL output from the run.
• src/mastra/workflows/main.ts — full pipeline: embed artifact → generate perspectives → model → analyze → build fragments → merge.
• src/mastra/workflows/fragments.ts — fragment-only workflow.
• src/mastra/agents/ — staged agents for perceiving, modelling, analyzing, building, and merging.
• src/viewer/ — local fragment viewer/server.
• .docs/proposal.md — original architecture proposal and category-theoretic design notes.
• docs/symbolic-memory-notes.md — notes on how curated/cleaned ontology views could ground symbolic memory.

Commands

Install dependencies:

bun install

Run the Mastra/Convex/viewer dev stack:

bun run dev

Run only the fragment viewer:

bun run viewer

Summarize the committed OWL artifact:

python3 scripts/summarize_owl.py unified.owl

Typecheck:

bunx tsc --noEmit

Portfolio artifacts

Portfolio images and extracted source-thread artifacts are currently generated in the recruiting workspace rather than committed here:

/Users/jesselee/dev/dd/recruiting/narrative/assets/ocra/

The useful page assets are:

• curated ontology map — readable view of the Buddhist/Zen semantic structure
• full OWL hairball — dense WebVOWL-style proof of the unified artifact
• source thread excerpt/full transcript — grounding input for the retrospective writeup

Interpretation

The dense unified.owl graph is good evidence that the prototype can decompose a large philosophical conversation into symbolic structure. The curated/minimal graph is probably more useful for downstream memory: it exposes the stable conceptual spine while hiding low-value assertion noise.