RuleMemory — Qwen Cloud MemoryAgent

▶ Demo video: https://youtu.be/r7dl2qs1vlI · ▶ Repo: https://github.com/cjw0076/rulememory-qwen-agent

A Qwen-powered MemoryAgent that remembers contest rules, deadlines, source evidence, and stale assumptions across sessions — and resolves contradictions instead of silently piling them up. Built for the Global AI Hackathon Series with Qwen Cloud, MemoryAgent track.

What makes it prize-caliber

Cross-session memory with provenance. Every remembered fact carries an ID, a source reference, and a stale-after window. Memory is a queryable graph, not a chat log.
Conflict resolution by supersession. When a new fact targets the same topic as an existing one (e.g. "use Python 2" → "use Python 3", or an updated deadline), the older fact is marked superseded — so memory never quietly holds a contradiction. Shown live in the demo.
Stale-assumption detection. Facts past their freshness window are flagged automatically, so the agent warns you before you act on a decayed assumption.
Grounded answers that cite their sources. Questions are answered over the remembered facts, and the answer points back at the specific entry IDs it used.
Qwen Cloud reasoning route. Fact extraction and answer synthesis route to the Qwen Cloud MaaS OpenAI-compatible endpoint (qwen-plus) when a key is configured; a captured real Qwen run is bundled as eval evidence (eval/SAMPLE_RUN_qwen_local.txt). The memory features above are model-independent and run in the live web console regardless.
A real web console + eval harness + tests. Ingest, ask, recall, a live memory table, and the full multi-step transcript — plus eval/demo.py and a passing test suite.

How it runs

contest rules / deadlines / operator notes
  → Qwen extracts structured facts (id + source + stale-after)
  → store.upsert into the RuleMemory graph
  → supersede any conflicting prior fact
  → flag facts past their stale-after window
  → Qwen answers questions grounded on remembered facts, citing entry IDs