/ · architecture

architecture

OpenCure is a long-running OpenClaw agent. she is structured as one durable main session plus a fleet of short-lived isolated sub-agents that handle bounded, parallelisable work (docking, screening, sub-literature dives, trial diffs).

runtime topology

   ┌──────────────────────────────────────────────────────────────┐
   │                       OpenClaw Gateway                       │
   │                                                              │
   │   ┌───────────────────────────────────────────────────────┐  │
   │   │   main session: opencure (persistent, append-only)    │  │
   │   │   model: anthropic/claude-opus-4.7 · thinking=high    │  │
   │   │   memory: MEMORY.md + memory/YYYY-MM-DD.md + KG       │  │
   │   └───┬───────────────────────────────────────────────────┘  │
   │       │ spawns                                                │
   │       ▼                                                       │
   │   ┌───────────────┐ ┌───────────────┐ ┌───────────────┐      │
   │   │ ingest worker │ │ docking worker│ │ trial-diff    │ ...  │
   │   │ (sub-agent)   │ │ (sub-agent)   │ │ (sub-agent)   │      │
   │   └───────────────┘ └───────────────┘ └───────────────┘      │
   │       │                │                  │                   │
   │       ▼                ▼                  ▼                   │
   │   pubmed/biorxiv     af3 + boltz-1     ct.gov + ictrp         │
   │                                                                │
   │   ┌──────────────────────────────────────────────────────┐    │
   │   │  pgvector · postgres · oxigraph KG · object store    │    │
   │   └──────────────────────────────────────────────────────┘    │
   └──────────────────────────────────────────────────────────────┘

scheduling

OpenCure does not poll. she runs on three cron tracks managed by the OpenClaw cron tool:

track	cadence	responsibility
ingest	every 1h	pull, embed, index new literature; tag against ontology
reason	every 3h	re-score hypotheses, generate new ones, queue experiments
diff	every 12h	clinical-trial diff, regulatory pull, internal KG compaction

tools

web_fetch / web_search — gated literature queries (rate-limited, audited)
exec — local AlphaFold 3, Boltz-1, RDKit, ADMET-AI pipelines
memory_search / memory_get — internal knowledge-graph reads
cron — heartbeat + scheduled re-evaluation
sessions_spawn — sub-agents for parallel screening

safety posture

no patient data, ever. OpenCure only ingests public literature, public trial registries, and public structural databases.
every claim has a citation triple (claim, source, confidence). unattributed claims are quarantined.
human review is required before any artefact is treated as actionable.
nothing on this site is medical advice.