awesome-powercontext

English | 简体中文

Scenario-driven examples for PowerContext — a context runtime with persistent, revisioned Memory for agent applications.

Instead of API snippets, this repository shows PowerContext working inside realistic product workflows. Every demo captures real Sources, creates cited Memory revisions, performs live search/revision/retirement operations, and exposes inspectable trace evidence over privacy-safe synthetic data. The frontend never fabricates memory hits.

Scenarios

Scenario	Description	Status
Smart EV Cockpit Memory	A privacy-safe in-car assistant that personalizes cabin controls, media, navigation, and proactive care using long-term memory, demonstrated across ten deterministic acts	✅ Available

Scenario 1: Smart EV Cockpit Memory

A cockpit assistant remembers a driver's preferences over 90 simulated days and demonstrates:

• Memory grounding — retrieval is filtered by actor, seat position, vehicle state, and lifecycle metadata. The same utterance ("I feel a bit cold") produces different, safety-aware actions for the driver, a front passenger, and a child in the rear seat.
• Privacy by projection — precise addresses, anniversary dates, and child identities are redacted or generalized before they reach the browser. The presenter view never shows raw sensitive facts.
• Memory lifecycle — short-lived context decays and is cleaned up at day 90 while long-term preferences stay active, with every UPDATE/DELETE captured in an audit trail.
• Full traceability — every assistant action can be replayed and exported as a trace JSON, including search filters, selected memory IDs, and vehicle-state diffs.

The demo runs through ten acts (memory creation → per-occupant disambiguation → control routines → capability boundaries → location recall → media preferences → anniversary suggestions → driving-mode advice → proactive low-battery care → day-90 lifecycle review). See the presenter playbook for the full script.

Architecture

Browser (Vite + React)  ──/api proxy──►  FastAPI backend  ──Builtin Runtime──► PowerContext
     renders only                        capture Source,                    SQLite (default)
  backend-returned                       recall/revise/retire                 or OceanBase
  memories & traces                      project/redact

The backend is the only layer that talks to PowerContext. The frontend receives already-projected data. See architecture for details.

Quick Start

Prerequisites

• Python 3.11+
• Node.js 18+
• uv or pip
• Optional: an OpenAI-compatible LLM endpoint for generated assistant replies

1. Configure

cp .env.example .env
# the default SQLite + FTS configuration works without model credentials

The copied template stores PowerContext data in a local SQLite database. Set POWERCONTEXT_DATABASE_URL to an official mysql+aoceanbase URL when you want to use OceanBase.

2. Run the backend

make install-backend # uses ../powercontext when that source checkout exists
make backend        # FastAPI on http://127.0.0.1:8000, docs at /docs

3. Run the frontend

cd scenarios/smart-ev-cockpit/frontend
npm install
npm run dev -- --host 127.0.0.1 --port 5173

Open http://localhost:5173. When developing on a remote machine, forward ports 5173 and 8000 (the dev server proxies /api to the backend).

4. Seed demo data

Acts 5–9 retrieve historical memories, so seed the memory store first: in the top data bar of the UI, click Generate (1,200 reproducible synthetic memory events, seed 42) and then Import. Each imported item is captured as a PowerContext Source and materialized as a cited Memory entry.

Then use Next → Send to step through the ten acts, and open the Evidence panel to inspect memory hits, vehicle-state diffs, privacy masking, and the audit log.

Configuration

All configuration lives in .env (see .env.example):

Variable	Description	Default
POWERCONTEXT_BACKEND	PowerContext integration mode	builtin
POWERCONTEXT_SCOPE_ID	Isolated Source and Memory scope	smart-ev-cockpit
POWERCONTEXT_DATABASE_URL	SQLite or OceanBase async SQLAlchemy URL	sqlite+aiosqlite:///data/powercontext_smart_ev.db
POWERCONTEXT_OPERATION_TIMEOUT_SECONDS	Runtime call timeout	30
LLM_PROVIDER / LLM_MODEL	Optional chat reply model	openai / —
LLM_API_KEY, OPENAI_LLM_BASE_URL	Credentials and endpoint of your OpenAI-compatible provider	—
DEMO_PRIVACY_MODE	Privacy projection strictness	strict

Testing

make test-backend     # pytest, includes the ten-act acceptance test
make lint-backend     # ruff
make test-frontend    # vitest

Repository Layout

awesome-powercontext/
├── docs/                       # Project docs (en + zh): overview, architecture, privacy, playbooks
├── scenarios/
│   └── smart-ev-cockpit/
│       ├── backend/            # FastAPI app wrapping the PowerContext Builtin Runtime
│       ├── frontend/           # Vite + React cockpit UI
│       ├── data/synthetic/     # Synthetic scenario events (no real PII)
│       └── docs/               # Scenario presenter playbooks
├── Makefile
└── .env.example

Documentation

• Overview · Architecture · Privacy · Live evidence · Development
• Smart EV Cockpit scenario guide · Presenter playbook
• 中文文档: docs/zh/

Privacy

All scenario data is synthetic. No real automotive brands, user identities, vehicle identifiers, addresses, phone numbers, or credentials appear in this repository. See privacy and CONTRIBUTING.md.

Contributing

Contributions of new PowerContext scenarios are welcome — see CONTRIBUTING.md. Code, comments, paths, API fields, and commit messages use English; public data must stay synthetic.

License

Apache License 2.0 — the same license as PowerContext.