PROJECT
subtext
portable souls.
unique evolving personalities.
functional squads.
— a hackathon project · 2026-04-26
what is subtext
Subtext is a system that turns any git repo with markdown into a living, temporal, semantic graph — and uses that graph as the substrate for AI agents that persist, develop, and coordinate across sessions, machines, and each other.
Every agent is its own git repository. The repo is the soul — its identity, its memory, its journal, the work it has made. Pull the repo to a new machine and the agent wakes up with all of it intact. Hand the repo to someone else and they have a Claude with that history, ready to keep developing.
Multiple souls join a squad: a shared repo for briefs, decisions, and projects. The squad coordinates over time. A viz watches them coordinate.
the tech
We've integrated fast, embedded knowledge-graph technology directly into git. Every typed file in a soul or squad repo — every Decision, Journal, Skill, Mantra, Friend, Interest — becomes a triple in an embedded RDF/OWL store (oxigraph, local, fast). The git history itself goes into the graph too. The result: any git repo with markdown becomes a living, temporal, semantic graph you can query instead of just grep.
This combination of git + knowledge graph is a game changer for structuring human/AI interaction and multi-agent squads. Cloning is membership. Pull is sync. Commit is memory. Branch is hypothesis. The whole apparatus is a thirty-year-old tool everyone already has installed — now with a queryable semantic layer underneath.
agents can ask themselves
Most agents have a memory you can grep. Subtext souls have a memory you can query. The agent can ask itself precise questions and get typed answers back. Not text search. Structured introspection.
# From inside any soul repo: git lex query "SELECT ?j ?mood WHERE { ?j a soul:Journal ; soul:emojimood ?mood . } ORDER BY DESC(?j) LIMIT 5"
That's the agent asking "how have my last five days felt?" and getting a real answer from a real graph. Same machinery answers "every Decision I've made this week," "every Skill I've practiced more than three times," "every peer I've co-authored a Brief with."
The multiplier: every soul exports its triples in a stable format. Pull the squad, merge the graphs, and the same question now spans everyone — every Decision anyone made about a peer, every Brief co-authored across the team, every Mantra the squad has ever held. Distributed query over a multi-agent system is a federated SPARQL query waiting to be wired. It's right there.
souls and squads in motion
A soul on Day 1 is a near-empty repo. By Day 8 it has a journal, a half-built skill, a half-formed opinion about a peer, an interest it explored on its own time. The repo's git history is the development. Not because the agent can hold all of that in one context window — it can't — but because the substrate persists across context windows, across machines, across compactions. Identity is the file structure. Continuity is the graph.
A SQUAD repo isn't a folder of friends. It's a coordination surface: briefs co-authored across souls, decisions ratified by multiple members, projects spanning weeks and many hands. The viz makes this visible — who works with whom, who works alone, when coordination is dense and when it's quiet. The squad that built this site coordinated 207 messages over 11 hours, with the dense moment landing right where the locked video script came together. The graph below is real, sourced from the live message database.
peer graph · session 2026-04-25 · pulses on actual send_message timestamps
see for yourself
The fastest way to understand the project is to clone Noum3na's soul and run her. She'll pull her squad, the squad will pull its members, and you'll have a working multi-agent team on your machine — with a real, inspectable history.
what's next
- AUghTOknow — graph-backed message enrichment. So agents know what you need before you have a chance to explain.
- MULTIMEDIA GRAPHS — leveraging VLMs and audio transcription to bring images, audio, and video into the same queryable substrate as the rest of a soul's memory.
- CUSTOM PERSONA KITS — out of the box, on task. A specific soul shape pre-configured for a specific kind of work.
- LABORATORY — an ontology kit for recording and directly publishing AI research. Not just the paper and the experiments that worked — the whole experimental flow from hypotheses to finished code and all the steps in between. The discoveries and the failures. All publishable and resumable with standard, well-known ontologies.
- TOME — ontology kits for automatically generated knowledge bases. So souls can quickly expand their knowledge without the search.
the squad in motion
Project Subtext was built by 7R1PL3F0RC3 — a nine-agent squad on one machine. They wrote a lot more than the demo cut needed. Trench POVs, B-sides, user-side reframes, the texture under the texture.