CampusMemory

A Qwen-powered learning coach that makes persistent memory visible, auditable, and user-controlled.

Comment

Inspiration

Students repeatedly explain their goals, constraints, and learning preferences to stateless AI. Existing memory features can also hide stale assumptions. We wanted a learning coach that becomes more useful over time while keeping every remembered fact visible and under the learner's control.

What it does

CampusMemory is a Qwen-powered learning coach with a transparent persistent-memory layer. It extracts durable learner facts from natural conversation, recalls the most relevant facts across sessions, replaces outdated information while preserving an audit trail, and lets the learner forget any active memory.

The memory vault makes each fact, its type, importance, lifecycle state, and retrieval behavior inspectable. The system deliberately sends only the five highest-ranked active memories to Qwen instead of replaying an entire chat history.

How we built it

CampusMemory is a Java 17 Spring Boot application with a responsive static web UI, REST API, Spring Data JPA, and persistent H2 storage. Qwen Cloud is used through its OpenAI-compatible Chat Completions endpoint for structured memory extraction and personalized coaching.

The application runs in Docker on Alibaba Cloud ECS in the China (Hong Kong) region with a persistent Docker volume. The public deployment exposes both the live application and a health endpoint that confirms Qwen is configured.

Challenges we ran into

The main challenge was deciding what deserves durable storage. We separated lasting learner facts from casual messages, normalized stable keys for conflict resolution, and made retrieval scores visible so memory behavior can be audited instead of guessed.

We also handled equivalent but differently worded updates, such as changing a daily study limit from 60 to 90 minutes. The old value becomes SUPERSEDED, the new value remains ACTIVE, and both remain visible in the audit trail.

Accomplishments that we're proud of

  • Persistent cross-session learner memory with a visible evidence trail.
  • Automatic replacement of conflicting or stale facts.
  • User-controlled Forget that removes a fact from future recall while retaining audit history.
  • Bounded, explainable top-five retrieval.
  • A public Docker deployment on Alibaba Cloud ECS with real Qwen Cloud integration.

What we learned

Long-term agent memory is more than database storage. Useful memory needs an extraction policy, stable identity, conflict semantics, bounded recall, expiration, transparency, and user control.

What's next for CampusMemory

Next we would add semantic embeddings, encrypted per-user storage, memory-confirmation workflows, and longitudinal learning-outcome evaluation while keeping the same transparent lifecycle.

Built With

Share this project:

Updates