What it does
Mimir MemoryAgent solves the fundamental problem of AI amnesia. Every LLM conversation starts from zero — no memory of past projects, preferences, decisions, or corrections. Mimir fixes this by wrapping a production-grade persistent memory backend (27 MCP tools, FTS5+vector hybrid search, AES-256-GCM encryption, Ebbinghaus decay) with Qwen Cloud reasoning models.
The agent demonstrates:
- Persistent memory across sessions — remembers projects, facts, preferences
- Cross-session context accumulation — gets smarter as memory grows
- Timely forgetting — Ebbinghaus decay removes stale/unused memories
- Critical memory recall — retrieves relevant context without dumping everything
- Human-in-the-loop surfacing — surfaces what it remembers at decision points
How we built it
- Memory backend: Mimir (Rust) — production persistent memory with SQLite, FTS5, vector search, and Ebbinghaus decay
- Reasoning: Qwen Cloud via OpenAI-compatible API (qwen-max model)
- Orchestration: Python agent loop: recall → reason → remember → groom
- Architecture: 3-layer — MemoryAgent Python orchestrator, MimirBridge for memory operations, QwenClient for reasoning
Why Qwen Cloud
Qwen Cloud's OpenAI-compatible API made integration seamless. We used qwen-max for its strong reasoning capabilities — the agent needs to both recall relevant context AND extract new facts from responses. The 1M-token context window on qwen-max-longcontext enables rich memory context without truncation, which is critical for the MemoryAgent track's cross-session requirements.
Challenges we ran into
- Qwen Cloud's hackathon voucher system requires separate application — built the agent to work with any OpenAI-compatible endpoint as fallback
- Mimir's MCP stdio protocol requires persistent process management — used direct CLI subprocess calls for hackathon simplicity
- Memory extraction (identifying what to store) is itself an LLM task — used a second Qwen Cloud call with a structured extraction prompt
Accomplishments
- Full agent loop working end-to-end in under 4 hours
- Demo shows clear progression across 5 sessions: introduction → recall → preferences → decay → decision support
- Architecture diagram and documentation in the repo are ready for production use
- Not a toy — Mimir is a real memory system with 27 MCP tools, used in production agents
What we learned
- Persistent memory transforms the agent experience — users don't need to repeat themselves
- Ebbinghaus decay is more intuitive than LRU for demonstrating "natural forgetting"
- Structured entity model (category, key, body_json) is the right abstraction for agent memory
- Qwen Cloud's API is reliable and fast for the recall→reason→remember loop
What's next
- Web dashboard for memory browsing and graph visualization
- Multi-agent shared memory pools (Agent Society track)
- LLM-based memory compression for long-term storage
- Cross-workspace memory federation
