Cadence

Deep-memory vertical agents for healthcare — pure Python, fully local, zero agentic frameworks.

Disclaimer: Educational / research prototype only. Never use for real medical decisions. Always consult qualified clinicians.

Live Demo

https://cadence-healthcare.vercel.app/

Status: Public UI is a client-side companion + share-code clinician brief. Live chat uses Groq llama-3.3-70b-versatile (env OPENAI_API_KEY on Vercel). If the key is missing or Groq errors, the badge switches to demo fallback. Full multi-layer memory + consent grants run locally (python run_patient.py / python run_clinician.py + Ollama). Do not enter real PHI.

Screenshots

Overview	Companion

Share code	Clinician brief

---

What it is

Vertical AI agents that remember — symptoms, history, preferences — across sessions.

• Multi-layer deep memory (session · episodic · semantic · knowledge · insights)
• Pure Python only (no LangChain, CrewAI, AutoGen, Mem0…)
• Fully local & free (Ollama + SQLite + sentence-transformers)
• Consent-scoped clinician brief / note draft
• Patient journey first: Baseline → Triage → Visit Prep → Care → Pattern → Recovery

---

Tech stack

Layer	Choice
Agents	Pure Python (no LangChain / CrewAI / Mem0)
Local LLM	Ollama (llama3.1)
Live UI chat	Groq llama-3.3-70b-versatile via Vercel /api/chat
Memory	SQLite + sentence-transformers (injectable embedder)
Consent	Scope-gated grants + audit table
UI	Static HTML / Tailwind CDN on Vercel

---

Quick Start

# 1. Ollama
ollama pull llama3.1

# 2. Python
python -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate
pip install -r requirements.txt

# 3a. Patient journey (recommended)
python run_patient.py

# 3b. Single triage agent
python run_agent.py

# 3c. Clinician grant / brief / note
python run_clinician.py grant --patient demo --clinician dr_lee --hours 48

Force a stage:

python run_patient.py --stage VISIT_PREP

In-session: type /stage CARE to switch.

Memory lives in data/ and survives restarts (gitignored).

Tests

pip install -r requirements-dev.txt
python -m pytest -q
npm ci && npx playwright install chromium && npm run test:e2e

---

Patient journey stages

Stage	Agent	Role
BASELINE	Baseline	Profile, allergies, meds, goals
TRIAGE	Triage	Symptom structure + cautious red flags
VISIT_PREP	VisitPrep	Questions + brief for the clinician visit
CARE	CareCompanion	Adherence, side effects, care-plan tasks
PATTERN	Pattern	Hypothesis correlations from memory
RECOVERY	Recovery	Milestones and “what better looks like”

Spec: docs/PATIENT_JOURNEY.md

---

Architecture

Memory Layers

1. Session / Working – recent turns
2. Episodic – timestamped events, symptoms, visits
3. Semantic – vector long-term facts
4. Knowledge – local RAG over guidelines
5. Insights – synthesized patterns (human-verified)

---

Project Structure

healthcare-deep-memory-agents/
├── docs/screenshots/        ← product screenshots
├── run_patient.py           ← patient journey CLI
├── run_clinician.py         ← grant / brief / note CLI
├── src/memory/              ← DeepMemory + ConsentStore
├── src/agents/
├── web/                     ← Cadence UI (Vercel)
├── tests/                   ← pytest (no torch / Ollama)
├── e2e/                     ← Playwright smokes
└── data/                    ← local DB (gitignored)

---

Security

See SECURITY.md. Educational prototype. Public chat messages go to Groq when live mode is on.

---

License

MIT (code). Any medical content you add keeps its original license.