feat: AP Memory Agent demo with Loom video

demo/ap-memory-agent/.env.example

@@ -0,0 +1,5 @@
+ANTHROPIC_API_KEY=your_anthropic_key
+EVERMEM_BASE_URL=http://localhost:1995
+EVERMEM_USER_ID=ap_agent_system
+# Required for seeded vendor history (matches ap_agent_system's group)
+EVERMEM_GROUP_ID=eb6618c4d52d3bf9_group

demo/ap-memory-agent/README.md

@@ -0,0 +1,126 @@
+# AP Memory Agent
+
+Accounts Payable automation agent built with **LangGraph + EverMemOS + Claude** for the [Memory Genesis Competition 2026](https://luma.com/n88icl03?tk=aawakR).
+
+**Core demo moment:** An agent catches a fraudulent or duplicate invoice because it remembers what happened weeks ago.
+
+## Demo Video
+
+<a href="https://www.loom.com/share/9834fe7b5fca45c78c304f7101b15ac6">
+  <img src="https://cdn.loom.com/sessions/thumbnails/9834fe7b5fca45c78c304f7101b15ac6-31131ab95b0f7652.jpg" width="600" />
+</a>
+
+[Watch the full demo on Loom](https://www.loom.com/share/9834fe7b5fca45c78c304f7101b15ac6)
+
+## Architecture
+
+```
+Streamlit UI (invoice form)
+        ↓
+LangGraph Orchestrator
+    ├── Node 1: invoice_agent     → normalise invoice, write episodic_memory, fetch vendor context
+    ├── Node 2: risk_agent         → Claude reasons over invoice + memory context, returns risk flags
+    ├── Node 3: approval_agent     → Claude makes approve/hold/reject decision with reasoning
+    └── Node 4: memory_updater     → write resolution to episodic_memory, update profile if risk ≥ 40
+                ↕                        ↕                    ↕                       ↕
+                            EverMemOS (Shared Memory Bus)
+```
+
+### Memory Strategy
+
+- **Invoice Agent:** Writes each invoice event as episodic memory; searches episodic + event_log (agentic retrieval); fetches profile separately; merges into `memory_context`.
+- **Memory Updater:** Always writes resolution as episodic memory; conditionally writes profile when risk_score ≥ 40 or decision is hold/reject
+- **Profile** is not searchable — fetched via GET /memories and merged manually
+
+## Setup
+
+1. **Prerequisites**
+   - Python 3.11+
+   - EverMemOS running locally (default: `http://localhost:1995`)
+
+2. **Install dependencies**
+   ```bash
+   pip install -r requirements.txt
+   ```
+
+3. **Configure environment**
+   ```bash
+   cp .env.example .env
+   # Edit .env with your ANTHROPIC_API_KEY
+   # EVERMEM_GROUP_ID is required for seeded vendor history
+   ```
+
+4. **Seed vendor history**
+   ```bash
+   cd demo/ap-memory-agent
+   python seed_memory.py
+   # Or: python seed_memory.py --clear   # Clear and re-seed
+   ```
+   From EverMemOS root: `
+   `
+
+   **Extraction timing:** EverMemOS uses boundary detection — messages are extracted into episodic memories when a "conversation episode" is complete. The seed sends user+assistant pairs with `sync_mode` to trigger extraction. Expect 1–2 minutes for the full seed. Episodic memories appear after boundary detection runs.
+
+5. **Run the UI**
+   ```bash
+   cd demo/ap-memory-agent
+   streamlit run streamlit_app.py
+   ```
+   From EverMemOS root: `uv run streamlit run demo/ap-memory-agent/streamlit_app.py`
+
+## Demo Script (for judges)
+
+1. Run `seed_memory.py` to populate vendor history
+2. Open Streamlit UI
+3. **Acme Corp** — Submit invoice `INV-2025-0042` for $500 → should **APPROVE** (clean history)
+4. **Globex Supplies** — Submit invoice `INV-2025-0105` for $1,200 → should **HOLD** (prior dispute)
+5. **Shadow LLC** — Submit invoice `INV-777` for $3,000 → should **REJECT** (duplicate caught by memory)
+
+## Test Examples
+
+### Streamlit UI (manual form)
+
+| Test | Vendor Name | Invoice Number | Amount | Expected |
+|------|-------------|----------------|--------|----------|
+| 1 | Acme Corp | INV-2025-0042 | 500 | APPROVE |
+| 2 | Globex Supplies | INV-2025-0105 | 1200 | HOLD |
+| 3 | Shadow LLC | INV-777 | 3000 | REJECT |
+| 4 | Acme Corp | INV-2025-0099 | 550 | APPROVE |
+
+### CLI test script
+
+```bash
+uv run python demo/ap-memory-agent/test_demo.py
+```
+
+Runs all 4 test invoices through the graph and prints decisions.
+
+## Project Structure
+
+```
+ap-memory-agent/
+├── .env                    # API keys (never commit)
+├── .env.example            # Template for env vars
+├── requirements.txt       # Dependencies
+├── ap_agent_graph.py      # LangGraph pipeline (4 nodes)
+├── evermem_client.py      # EverMemOS API wrapper
+├── seed_memory.py         # Pre-populate vendor history
+├── streamlit_app.py       # Streamlit UI
+└── README.md
+```
+
+## Environment Variables
+
+| Variable | Description | Default |
+|----------|-------------|---------|
+| `ANTHROPIC_API_KEY` | Claude API key | required |
+| `EVERMEM_BASE_URL` | EverMemOS URL | `http://localhost:1995` |
+| `EVERMEM_USER_ID` | Memory scope | `ap_agent_system` |
+| `EVERMEM_GROUP_ID` | Group for seeded data | `eb6618c4d52d3bf9_group` |
+
+## EverMemOS API Notes
+
+- **POST /api/v1/memories** — Store message (MemorizeMessageRequest: message_id, create_time, sender, content, role)
+- **GET /api/v1/memories** — Fetch by memory_type (profile, episodic_memory, etc.)
+- **GET /api/v1/memories/search** — Search episodic_memory, event_log, foresight (profile NOT supported)
+- **retrieve_method="agentic"** — Use for agent queries (LLM-guided retrieval)