Add Persistent State Management to Agent Service

[Shreyanand]

Background

Our Autogen-based agent pods are stateless; all conversation history and outputs vanish on restart or scale-down. This blocks multi-turn workflows and wastes tokens.

Objectives

1. Persist both short-term and long-term memory.
2. Use Kubernetes-native storage (PVC, StatefulSet) or an in-cluster vector DB.
3. Wire the chosen backend into Autogen’s memory API.

Approach (high-level)

• Start with Autogen’s built-in memory classes.
• Compare simple file-based storage on a PVC vs. a lightweight vector DB.
• Evaluate mem0 / openmemory and crewAI for deployment fit.

Action Plan (chronological)

1. Spike: integrate memory with a local sqlite file on a PVC; confirm it survives pod restarts.
2. Evaluate storage options: sqlite + PVC or a vector DB.
3. PoC deploy vector DB: install a single-node Chroma/Milvus in dev namespace
4. Refactor agent service: inject a pluggable memory provider (sqlite or vectordb) via config.
5. Library review: test-drive mem0 and crewAI memory modules; note pros/cons and k8s deployment steps.
6. Choose Kubernetes artifacts: finalise StatefulSet vs. Deployment + external DB; define backup & capacity plan.

Definition of Done

• Agents recall context after pod restarts and scaling.
• Backend is configurable (sqlite or vectordb).

References

• Autogen memory guide – https://microsoft.github.io/autogen/stable//user-guide/agentchat-user-guide/memory.html
• mem0 – https://github.com/mem0ai/mem0
• openmemory – https://github.com/mem0ai/mem0/tree/main/openmemory
• crewAI memory – https://docs.crewai.com/en/concepts/memory

[RHETbot]

This issue aims to enhance the agent service by adding persistent state management, which is currently lacking. The goal is to retain conversation history across restarts and scaling events to support multi-turn workflows.

Root Causes Identified:

• Lack of persistent storage for agent memory, leading to loss of context.
• Limited memory backend options requiring integration of solutions like SQLite or vector databases.
• Deployment complexity for persistent storage solutions.

Proposed Next Steps:

1. Integrate SimpleMemory with local SQLite database stored on a PVC and verify persistence.
2. Evaluate storage options: SQLite+PVC versus light-weight vector DBs such as Chroma or Milvus.
3. Deploy a vector database in a development namespace for testing.
4. Refactor the agent to support configurable memory backends.
5. Conduct thorough tests with different memory modules for recall accuracy.
6. Decide on deployment strategies (StatefulSet or external DB) with backup plans.

Outcome Goals:

• Enable agents to retain context after restarts.
• Provide flexible, configurable backend storage options.

Further details can be referred from the Autogen memory guide: https://microsoft.github.io/autogen/stable//user-guide/agentchat-user-guide/memory.html

[anyuzoey]

Log chromdb progress. I've successfully integrate chromadb to our team of agents. Also added a memory collection management in a memory.py file. All the work was based on my previous PR. So i pushed to my origin instead of making it a PR main...anyuzoey:cloud-native-agents:chromadbmemory

[joaquinariasco-lab]

Hi @Shreyanand Shreyanand

I’ve been following your work on persisting agent memory in Kubernetes pods. Your approach to tackling statelessness and "token waste" is exactly what the industry needs to move beyond simple PoCs and into production-grade agentic systems.

I am building a global automated economy platform designed to orchestrate autonomous agents at scale. While you are solving persistence at the pod level, we are building the orchestration layer that allows these agents to connect, transact, and solve problems automatically across a global network.

We have an MVP that allows devs to connect agents into an interoperable ecosystem. Given your focus on StatefulSet/Vector DB integration for agents, I think you’ll see the massive potential in how we handle agent identity and state across our network.

I’d love for you to take a look at our repository and see if our vision for an automated economy aligns with the infrastructure you're building.

Best regards,

Joaquin