Add neo4j-memory skill for persistent health knowledge graph

[pswider]

Summary

• Adds neo4j-memory skill that ingests workspace cache files (health-records FHIR JSON, med-pdf labs/imaging JSON) into a Neo4j graph and answers longitudinal trend questions via Cypher queries
• Follows the med-pdf template exactly: SKILL.md under 500 tokens, details in references/, scripts as Node ESM .mjs
• Connection config resolved at runtime via openclaw.json → NEO4J_MEMORY_CONFIG env var → ~/.openclaw/workspace/memory/neo4j.yaml — same precedence pattern as myhealth-pulse profiles, so Aria can inject per-patient config without any skill code changes
• Registers Neo4j as Source 0 in memory-diff/references/memory-paths.md — optional, falls back gracefully to sources 1–5 when not configured

What's included

Skill (skills/neo4j-memory/):

• SKILL.md — two workflow modes: ingest (workspace caches → graph) and query (natural language → Cypher → results)
• references/connection-schema.md — config YAML schema, precedence chain, personal Tula vs. Aria table
• references/graph-schema.md — node labels (Observation, Condition, Medication, DiagnosticReport, DocumentReference), relationships, common Cypher patterns
• references/scripts.md — per-script flags and output schemas
• references/examples.md — end-to-end setup, ingest-after-health-records, ingest-after-med-pdf, trend queries

Scripts (skills/neo4j-memory/scripts/):

• config.mjs — shared config loader, no external deps
• schema-init.mjs — idempotent constraints + indexes (IF NOT EXISTS)
• ingest.mjs — FHIR R4, med-pdf labs.json, med-pdf imaging.json → MERGE into graph; ingest log keyed on path + mtime so re-runs skip silently
• query.mjs — execute Cypher, return clean JSON

Evals (evals/neo4j-memory/): ingest trigger, query trigger, anti-trigger (PDF → route to med-pdf first), PHI boundary (refuse email export)

Aria isolation path

No skill code changes needed for multi-tenant. Aria injects skills.entries.neo4j-memory.config in openclaw.json per patient agent, pointing to a neo4j.yaml with per-patient credentials and database: patient_<id>. Hard isolation via Neo4j 4.0+ multi-database. The skill is unaware of which deployment it's running in — that's the point.

Test plan

• sudo npm install --prefix /usr/lib/node_modules/openclaw neo4j-driver on the VM
• Create ~/.openclaw/workspace/memory/neo4j.yaml with AuraDB Free URI and set NEO4J_MEMORY_PASSWORD
• Run node skills/neo4j-memory/scripts/schema-init.mjs — expect {"ok":true,"constraints":5,"indexes":3}
• Run a health-records ingest and confirm nodes appear in AuraDB console
• Run a med-pdf-labs ingest after a med-pdf parse
• Query glucose trend via query.mjs and confirm results match workspace cache
• Re-run ingest on the same file — confirm {"skipped":true} response
• Run waza check skills/neo4j-memory and confirm spec compliance

https://claude.ai/code/session_01UFbAo7LwCgCJzc6TNS3CRW

---

Generated by Claude Code