CampusMemory

A transparent persistent-memory learning coach powered by Qwen Cloud.

Live demo: http://47.239.40.162

Health check: http://47.239.40.162/api/health

Demo video: CampusMemory — Transparent Persistent Memory for Qwen Cloud

CampusMemory is a MemoryAgent track submission for the Qwen Cloud Global AI Hackathon. Unlike a stateless chatbot, it extracts durable learner facts, recalls only the most relevant memories in later sessions, replaces conflicting facts, expires short-lived context, and gives the learner a visible memory vault with a one-click forget control.

Why it matters

Students repeatedly explain their goals, schedule, skill level, and recurring mistakes to generic AI assistants. CampusMemory turns those facts into a compact, inspectable memory layer so every new session starts with useful context—without hiding what was stored.

Three-session demo

1. Session 1: I am learning Java backend, preparing for an internship, and can study 1 hour per day.
2. Click New session, then ask: Plan today's study. The UI shows the cross-session memories and their retrieval scores.
3. Say: I can now study 2 hours per day; the old limit is outdated. The stable daily_study_time key causes the one-hour memory to be superseded. The audit trail remains visible.
4. Use Forget to deactivate any active memory and prove that it is no longer recalled.

Memory lifecycle

• Extract: Qwen structured output converts chat into typed durable facts.
• Normalize: stable keys identify facts that can conflict, such as daily_study_time.
• Retrieve: a bounded top-5 context uses 55% importance + 30% keyword overlap + 15% recency.
• Replace: a newer value deactivates the previous value and links the audit trail.
• Expire: memories with a TTL are deactivated after expiresAt.
• Forget: the learner can deactivate an active memory from the UI.

Architecture

flowchart LR
    subgraph ECS["Alibaba Cloud ECS · Ubuntu 22.04"]
        subgraph D["Docker container"]
            W["Static web UI"] --> A["Spring Boot REST API"]
            A --> O["Memory lifecycle service"]
            O --> R["Bounded top-5 retrieval"]
        end
        R --> H[("Persistent H2<br/>Docker volume")]
    end
    U["Learner browser"] -->|"HTTP :80"| W
    O --> Q["Qwen Cloud API<br/>structured extraction + coaching"]

Loading

The Qwen API is stateless. CampusMemory owns the persistent memory lifecycle and sends only the five highest-scoring active memories in each prompt.

For submission forms and video editing, download the 1600×900 architecture PNG.

Run locally

Requirements: Java 17 and Maven 3.9+.

export DASHSCOPE_API_KEY="sk-your-key"
mvn spring-boot:run

Open http://localhost:8080. Without an API key, the app starts in clearly labelled DEMO MODE with a deterministic extractor, so reviewers can inspect the memory lifecycle. A judged Qwen Cloud run must set DASHSCOPE_API_KEY.

Configuration:

Variable	Default	Purpose
DASHSCOPE_API_KEY	empty	Qwen Cloud API key
QWEN_BASE_URL	https://dashscope-intl.aliyuncs.com/compatible-mode/v1	OpenAI-compatible Qwen endpoint
QWEN_MODEL	qwen3.7-plus	Qwen model ID
DATA_DIR	./data	Persistent H2 directory
PORT	8080	HTTP port

Run with Docker

export DASHSCOPE_API_KEY="sk-your-key"
docker compose up -d --build
curl http://localhost:8080/api/health

API

• POST /api/chat — extract, store, retrieve, and answer.
• GET /api/memories?userId=... — visible memory and audit trail.
• DELETE /api/memories/{id}?userId=... — learner-controlled forgetting.
• GET /api/health — deployment smoke test and Qwen configuration status.

Tests

mvn test

The integration tests verify conflicting-memory replacement, semantic-key alias repair, retained audit history, and manual forgetting removal from recall.

Alibaba Cloud deployment

The live judging deployment runs in Docker on Alibaba Cloud ECS in the China (Hong Kong) region. Administration uses Alibaba Cloud Session Manager, while the security group exposes only HTTP port 80 to the public internet.

• Public app: http://47.239.40.162
• Public health evidence: http://47.239.40.162/api/health
• Deployment instructions: deploy/README.md
• Alibaba Cloud proof code: deploy/aliyun-ecs-user-data.sh
• Detailed architecture: docs/ARCHITECTURE.md

Privacy and safety

• The extraction prompt rejects casual chat and sensitive secrets.
• API keys are environment variables and are never stored in the repository.
• Users can inspect and forget every active memory.
• Inactive memories stay in the MVP audit view; production deployments should add configurable hard deletion and encryption at rest.

License

MIT