Skip to Main Content
Talk Intermediate First Talk

No Reflection, No Insight: Why Open-Source AI Agents Need Meta-Memory

Proposal status is Rejected
Session Description

Open-source AI agents today can store and recall information using RAG, tools etc. But the problem is that they don’t know what they’ve forgotten.

They can’t check if a memory is missing, outdated, or wrong. They don’t reflect. This talk introduces a missing piece in agent design: meta-memory, which is memory about memory.

In 15 minutes, I’ll cover:

  • What meta-memory is and why open-source agents need it

  • Real examples of agents hallucinating or repeating due to missing memory checks

  • How to add basic meta-memory using Valkey, Postgres, Qdrant

  • A simple open-source pattern for confidence tracking and gap detection

We should move beyond recall and build agents that know what they don’t know, using the FOSS tools we already have.

Key Takeaways

  • The key limitation in today’s OSS agent memory

  • What meta-memory enables: confidence, awareness, fewer mistakes

  • How to build a basic meta-memory layer in your stack

  • A simple OSS blueprint you can implement today

References

https://github.com/archit15singh/meta-memory-agent-patterns

Session Categories

Technology architecture
Engineering practice - productivity, debugging

Which track are you applying for?

Main track

Speakers

Archit Singh
Archit Singh Senior Software Engineer

Archit Singh - Senior Software Engineer L5 | Deliveroo

Archit is a senior backend and infra engineer with 8+ years of experience building scalable systems across AI, distributed pipelines, and observability platforms. He has been part of engineering efforts at companies like multiple high growth startups, working on AI agents, GenAI infrastructure, healthcare pipelines, and real-time ingestion systems.

At heart, Archit is a systems thinker who blends open-source AI with cognitive systems and epistemology. He’s currently exploring stateful LLM architectures - rethinking memory not just as retrieval, but as reflection. He writes, builds, and experiments with Redis, Postgres, and Python to push the boundaries of open, trustworthy AI.

You can find him writing about infrastructure, cognition, and engineering at scale on LinkedIn.

https://www.linkedin.com/in/archit15singh/