Engram

A memory agent that learns workflows, gets faster at them, and self-corrects when they go stale.

Engram is a submission for the Global AI Hackathon Series with Qwen Cloud — Track 1: MemoryAgent.

It combines two memory mechanisms no existing agent demo tackles together:

• Procedural memory: watches a workflow once, distills it into a structured playbook, executes it faster and with fewer tokens each subsequent session
• The Forgetting Engine: a background consolidator that detects contradictions in playbooks, decays stale memories, and updates the playbook — so the agent doesn't just get faster, it stays correct as the world changes

Architecture

User
↓
Next.js Frontend (UI + API Routes)
↓
FastAPI Backend (Qwen-Agent framework)
├── Active Agent (Qwen3, non-thinking, low latency)
│ ├── Retrieves playbook from skills/ store
│ ├── Executes workflow steps via MCP tools
│ └── Writes episodic memories post-session
│
├── Consolidator (Qwen3, thinking-mode, async)
│ ├── Contradiction Detector
│ ├── Decay Scorer
│ ├── Belief Revision Engine
│ └── Memory Diff Logger
│
└── MCP Server (custom)
├── write_memory
├── update_memory
├── deprecate_memory
├── query_memory (semantic + BM25 + recency + importance)
└── get_playbook
↓
Memory Store (SQLite + ChromaDB)
├── Episodic memories
├── Procedural memories (playbooks)
└── Semantic memories
↓
Alibaba Cloud ECS

Setup

Prerequisites

• Python 3.10+
• Node.js 18+
• A DashScope API key (from Alibaba Cloud Model Studio)

Backend

cd backend
python -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env
# Add your DASHSCOPE_API_KEY to .env
uvicorn api.main:app --reload --port 8000

Frontend

cd frontend
npm install
cp .env.local.example .env.local
# NEXT_PUBLIC_API_URL=http://localhost:8000
npm run dev

Open http://localhost:3000

Tracks & Requirements

• Track: Track 1 — MemoryAgent
• Models: Qwen3-32B (active agent, non-thinking), Qwen3-32B (consolidator, thinking mode)
• Infrastructure: Alibaba Cloud ECS
• Memory: SQLite (structured) + ChromaDB (semantic vectors)
• Protocol: Custom MCP server for all memory operations

License

MIT — see LICENSE