Self-learning multi-agent RAG with layered memory: Qdrant + Hindsight architecture breakdown
Odd_Ranger5114 · reddit · 2026-09-30
A Reddit deep-dive describes a multi-agent architecture that makes RAG systems self-learning, fixing the flaw where standard RAG repeats the same errors on similar prompts.
Pipeline:
- Intake: classifies intent, urgency, and query scope
- Layered retrieval: static docs from Qdrant vector storage, past context/rules from Hindsight memory
- Draft & critique: a Critic Agent checks responses against known execution rules, capped at 2 redrafts to control cost/latency
- Async reflection: after answering, a Reflection Agent evaluates outcomes; at confidence ≥0.8 it converts the resolution into a permanent procedural rule
Four memory layers (instead of dumping raw transcripts into the vector store):
- Episodic: raw interaction histories
- Procedural: execution rules and learned constraints, shared by drafting and critic agents
- Semantic: infrastructure entities and facts
- Preference: user formatting and organizational preferences
Example: hitting an HTTP 429 during a large DB sync, the agent switches to smaller batches; the reflection worker then writes a rule "cap batch size at 50 for database sync tasks," instantly recalled next time.
Takeaways: keep execution rules separate from chat logs to avoid noisy retrieval; cap critique loops; run reflection asynchronously.
Related event: Hierarchical Memory Multi-Agent Architecture Makes RAG Self-Learning(2 posts)→
More from coding & agent
- Long-running cloud agents are now something you can run at home — jobergum · 2026-09-30
- Cloudflare cf CLI turns any webpage into Markdown in one command, keeping raw HTML out of agent context — dinasaur_404 · 2026-09-30
- A fixed agent workflow doesn't guarantee fixed output: path determinism ≠ result determinism — gethackteam · 2026-09-30
- One operator commanding seven AI agents: swarming orchestration demo called 'disturbing' — sethlazar · 2026-09-30
- DHH's AI coding workflow sparks debate: speed vs. understanding — georgemillo · 2026-09-30
- Delegance Turns a 1000-Page Insurance Binder into a Weaviate Knowledge Base in ~2 Minutes — CShorten30 · 2026-09-30