Idea
Add a background consolidation pass over our experiment memory, similar to OpenClaw's "dreaming" concept and the pattern OpenAI describes for ChatGPT memory.
Right now memory entries just accumulate. A periodic consolidation step could:
- Stage recent runs/signals without touching durable memory yet
- Cluster recurring themes/failure modes across experiments (e.g. same bug hit in multiple strategies)
- Score candidates (recall frequency, diversity of contexts, confidence) and only promote entries that clear a threshold
- Drop anything flagged untrusted or auto-generated noise before promotion
- Keep a human-readable log of what got promoted and why (a "dream diary" style summary), so it stays inspectable rather than a black box
This would help keep experiment memory from becoming a dumping ground of every run, while surfacing patterns (repeated failure causes, recurring good configs) that are easy to miss when just scrolling logs.
References
Open to bikeshedding on scope/whether this is worth the complexity vs. simpler manual curation.
Idea
Add a background consolidation pass over our experiment memory, similar to OpenClaw's "dreaming" concept and the pattern OpenAI describes for ChatGPT memory.
Right now memory entries just accumulate. A periodic consolidation step could:
This would help keep experiment memory from becoming a dumping ground of every run, while surfacing patterns (repeated failure causes, recurring good configs) that are easy to miss when just scrolling logs.
References
Open to bikeshedding on scope/whether this is worth the complexity vs. simpler manual curation.