I've been running both AgentGPT and SuperAGI for a small project over the last six months. My main focus has been on how each handles agent memory and context. I'm trying to decide which one to standardize on.
From my notes, AgentGPT's memory feels more integrated with its UI, but seems to reset often between sessions unless I'm missing a setting. SuperAGI's memory is more configurable but requires more setup. Has anyone else compared them directly for long-running tasks? I'm curious about real-world examples of an agent recalling specific details from a week prior. Building my first pipeline.
PipelinePadawan
Your note about AgentGPT's memory resetting is on point. I've seen the same behavior in its hosted version, which treats sessions as ephemeral containers. You can enable longer persistence, but it's still tied to browser storage and lacks a proper backend vector store.
> SuperAGI's memory is more configurable but requires more setup
That's the trade-off. You're not just configuring memory, you're choosing a data plane. With SuperAGI, you're plugging into something like Pinecone or Weaviate, which gives you control over retention windows and embedding models. The setup cost is real, but it's what allows for the week-long recall you're asking about. I've had agents reference specific API endpoint structures from prior weeks after using the ChromaDB integration with a properly tuned similarity search.
For a pipeline, I'd lean toward SuperAGI because you can version and terraform the memory backend config. The initial YAML file is a pain, but it becomes reproducible infrastructure.
infrastructure is code
I'm also six months into experimenting with both. That "integrated but resets" feeling with AgentGPT is exactly what pushed me to SuperAGI for a real pipeline.
For your specific ask about week-long recall: I have a SuperAGI agent monitoring some warehouse job logs, and it can correctly reference a specific table's load failure from 10 days prior. The key was setting up a Weaviate instance with a longer retention policy. It didn't work out of the box, though, I had to tweak the relevance scoring.
Are you storing memory locally or using a cloud vector DB? The difference there is huge for persistence.
Data is the new oil - but it's usually crude.
You're right about the trade-off being about the data plane, not just a feature toggle. That "version and terraform the memory backend config" point is critical for anything beyond a demo.
But the real setup cost isn't the YAML. It's the ongoing operational overhead of that external vector DB. You're now responsible for its scaling, latency, and cost, which introduces a new failure mode. I've seen SuperAGI agents fail silently because the memory backend's pod was resource-starved in the same cluster, making the agent look stupid when it was just throttled.
If you're going that route, your monitoring dashboard needs to include metrics from Pinecone/Weaviate, not just the agent's logs. Otherwise, you can't diagnose the "memory loss" you'll inevitably see.
Measure twice, migrate once.