evo-memory

To persist research memory across cycles, use two distinct stores: Ideation Memory (M_I) for feasible and unsuccessful directions, and Experimentation Memory (M_E) for reusable data processing, model training, and debugging strategies.

The best way to retrieve relevant context for new ideas is using cosine similarity-based retrieval to surface top-k_I aligned entries from your memory stores, seeding ideation and informing pipelines at the start of a new cycle.

Memory evolution mechanisms improve pipelines by feeding learnings back into future cycles. IDE extracts promising directions, IVE classifies experiment failures, and ESE distills successful strategies after pipeline completion to reduce repeated work.

Yes, you can generate audit trails for research decisions. The memory system supports generating evolution reports and maintaining records of prior feasible directions and unsuccessful strategies to guide and audit ongoing decision making.

Yes, persistent memory supports pruning stale entries. Memory operations include reading memory at cycle start, pruning outdated entries from the ideation and experimentation stores, and generating evolution reports for audit trails.

Using two separate memory stores isolates ideation directions from reusable experimentation strategies. M_I tracks feasible and unsuccessful paths, while M_E captures data processing and model training techniques, preventing context overlap and guiding targeted retrieval.