The author has developed a Rust project called SelMem to experiment with imperfect, path-dependent memory for large language models. This approach mimics human memory by allowing agents to compress, forget, and reconstruct memories over time rather than storing perfect copies of interactions.
- Two identical agents undergo the same initial interactions until one experiences a divergence.
- The system removes original context, relying on compressed and potentially distorted recalled memories for subsequent turns.
- Original inputs are stored separately for comparison but are not used during recall to test memory fidelity.
- The goal is to determine if differences in experience survive the selective memory process or produce interesting noise.
The author invites others working on LLM agents and memory to try breaking the idea, noting that while the utility is unproven, the experiment aims to explore how such mechanisms affect agent uniqueness.