evo-memory

What it does

Keeps a persistent learning layer across ideation and experimentation by storing historical directions and reusable strategies. It implements three evolution mechanisms: IDE to extract and store promising directions after ideation, IVE to classify why a failed method failed and update memory accordingly, and ESE to distill reusable data processing, model training, architecture, and debugging strategies after successful experiments. It provides templates and memory locations for Ideation Memory (M_I) and Experimentation Memory (M_E), plus procedures for loading prior knowledge at cycle start and writing evolution reports.

How it works

• Maintains two memory stores: Ideation Memory (M_I) and Experimentation Memory (M_E).
• IDE (Idea Direction Evolution): after research-ideation completes, reads M_I, runs an IDE prompt, abstracts directions to reusable levels, updates or appends M_I entries, prunes exhausted feasible directions, and writes an evolution report.
• IVE (Idea Validation Evolution): triggers after failed experiments or no executable code within budget; uses an IVE prompt to classify failure as FAILED(NoExecutableWithinBudget), FAILED(WorseThanBaseline), or NOT_FAILED; updates M_I accordingly, and provides a five-question diagnostic to distinguish implementation vs fundamental failure; escalation rules apply for repeated implementation failures.
• ESE (Experiment Strategy Evolution): after a successful experiment-pipeline, runs an ESE prompt to generate DATA SUMMARY and MODEL SUMMARY; populates M_E with data processing and model training strategies, then extracts architecture decisions and debugging patterns from trajectory logs, assesses generality, updates M_E, and writes an evolution report.
• Reading memory at cycle start: loads M_I and M_E, summarizes relevant entries, and injects them into the current context for ideation or experiments.
• Versioning and evolution reports are maintained under the memory framework; pruning rules apply to stale or superseded entries.

When to use it

• After completing a research-ideation cycle to update Ideation Memory (M_I).
• After completing or failing an experiment pipeline to update memory (M_E or M_I via IVE as applicable).
• When starting a new research cycle to load prior knowledge.
• When user asks about research memory, learned patterns, or cross-cycle knowledge.
• When user mentions evo-memory, update memory, research history, or evolution.

What it can touch

• Memory files: /memory/ideation-memory.md and /memory/experiment-memory.md.
• Templates and example injections described in the skill.
• It operates via prompts (IDE, IVE, ESE) and trajectory logs to extract reusable knowledge.

Caveats

• Do not use evo-memory for running experiments (use experiment-pipeline), debugging experiment code (use experiment-craft), or generating ideas (use research-ideation).
• Retrieval relies on embedding-like similarity and provided templates; applicability depends on cycle context and recorded metadata.
• Entries may be pruned if stale or superseded; version tracking and evolution reports are required for audit.