Memorii - a memory system for agents

I’ve been playing around with OpenClaw and Hermes since they were released. By May, I was frustrated: they kept forgetting basic facts and rules I’d repeated many times. I thought they needed a better memory system.

That became Memorii.

Before I jumped in, I went through about 25 papers on memory systems from the last few years (see references below). A few things stood out to me: having an ontology helps, keeping an execution state history helps, and maintaining a solver graph helps as the LLM searches for a solution.

This is exactly what I’m doing with Memorii. I’m building it in four stages.

Build the memory system: ontology, execution state, and solver graph.
Make LLMs manage the memory.
Explore periodic self-reflection as a learning signal for LLMs, including training on their own retrospections.
Build a system to update the KV cache at every layer using memory stored in Memorii.

I’ve completed step 1. Memorii now works with OpenClaw, Hermes, and Pi and tracks ontology, execution state, and a solver graph. Component-level benchmarks look encouraging. I’m still designing and running long-running agent-task benchmarks. I’ll share more with examples of how Memorii helps my agentic workflows.

References

These informed how I think about memory, task state, model integration, and evaluation.

Memorii - a memory system for agents

References

Agent memory and experience

Harnesses and task state

Model memory and KV caching

Memory evaluation