Qwen Cloud Hackathon — MemoryAgent Track

An agent that gets smarter every session.

Perseus Qwen Memory demonstrates persistent memory, timely forgetting, and cross-session compounding — the three pillars of the MemoryAgent track. Powered by Qwen Cloud + Mimir/Engram-rs.

Track Alignment

3/3

Persistent memory across sessions, timely forgetting via confidence decay, increasingly accurate over time.

Sessions

5 in demo

Five-session demo arc showing compounding accuracy. Same question, Session 1 vs Session 5 — graph included.

Memory Types

6 categories

Fact, decision, preference, lesson, insight, context. Each with its own decay curve and confidence model.

Track Requirements Checklist

The strength to double down on: every track requirement mapped to a concrete feature with a demo beat.

Requirement	How We Deliver	Demo Beat
Persistent memory across sessions	Mimir/Engram-rs SQLite backend. Memory stored in `mimir.db`, survives process restart.	"Session 1: I tell the agent my name. Session 3: it greets me by name without being told again."
Timely forgetting	Confidence decay per memory type. Un-reinforced facts lose confidence over configurable half-lives.	"An old fact about Pinecone at 15% confidence. The agent flags it as uncertain rather than asserting it."
Increasingly accurate	Cross-session compounding. Agent generates insights from patterns across sessions.	"Session 1: 60% accuracy on project questions. Session 5: 92% — same question set, compounding knowledge."

The Contradiction Beat

The most track-aligned demo moment. Old knowledge doesn't disappear — it degrades in confidence until contradicted by new facts.

Session 2:
  You: "We use Pinecone for vector search."
  Agent: Stored. [Pinecone → vector_search] confidence=0.90

Session 4:
  You: "We switched to pgvector last month."
  Agent: Updating — I had Pinecone at 40% confidence (un-reinforced for 2 sessions).
  Agent: Superseding with [pgvector → vector_search] confidence=0.90. Old fact deprecated.

Accuracy Curve

Same question set across 5 sessions. The graph you're told no other entrant will have.

Quickstart

run

# Set Qwen Cloud API key
export DASHSCOPE_API_KEY="your-key"

# Run the 5-session demo
python -m agent.main

# Switch to self-hosted Engram-rs
MEMORY_BACKEND=engram python -m agent.main

Qwen Memory on GitHub →