PI

Most agents forget you the moment the chat ends. Pi doesn't. Three-tier memory, an engineering loop that learns from its own bugs, now running on Qwen Cloud.

Comment

posted an update —

Pi is live on Qwen Cloud + Alibaba Cloud ECS

Pi started as a personal engineering-loop agent with three-tier memory. For this hackathon, I turned it into a genuinely Qwen-native memory agent:

  • Qwen leads every routing tier — via Alibaba Cloud Model Studio / DashScope (core/providers/qwen.py). It writes session summaries, distills durable facts, and generates the embeddings behind recall.
  • Hybrid retrieval — dense cosine over Qwen text-embedding-v3 fused with BM25, wired directly into the turn loop. Recall happens automatically, not just on request.
  • Real forgetting, not just storage — auto-expiry from phrasing, neglect decay, semantic dedup, and Qwen-adjudicated contradiction detection, with a forgetting ledger recording what was let go and why.
  • Two bugs fixed that actually matter for a memory agent: a sync job silently wiping local-only embedding/decay columns every 5 minutes, and forgotten rows being hard-deleted before they could be verified as zero-influence on recall. Both fixed with an archive-first approach.

Deployed and verified live on Alibaba Cloud ECS (Singapore) — backend running, health check confirmed via Qwen-routed response.

Repo's public: github.com/Ashar117/pi

Log in or sign up for Devpost to join the conversation.