[andai]

What's the point of the scratch pad? Isn't the same data already in the context? Or does it help because contexts are lossy and bias towards the start and end?

Similar question with the to-do list. Do they actually help task completion? Is there any research on that? I think they're less helpful with more recent models, but maybe they still help with smaller ones?

The system prompt asking it to make a plan before starting work does sound helpful though. (Of course it would also be great to see numbers there :)

[ELRayano]

Hi, There are few benefits from using scratchpad or any other external platform : - Be agnostic of the LLM you use, tomorow, if the prices of the llm you use are exploding, you can still reuse another LLM by pointing it the scratchpad repository you have. Then, modulo the level of verbosity you had on scratchpad (or other), you'll avoid lossing time ro reexplain everything to the new llm - You can avoid the "summarized" effect obtained through context compaction events . This effect makes accurate and so potential important information a bit more lurry (numbers turned into adjectives, etc/ Scratchpad or Obsidian or any other external solution you might imagine would act as "case fact blocks" that are a solution recommended to mitigate the above effect and thus make the accurate information still available. You can imagine a system where you ask your LLM to read some files within your external storage after each compaction for exemple with a hook or anything else.

Regarding the todolist, from my pov, it's just a basic principle of work segmentation and accuracy with some traceability. You are better when you can divide your work into chunks that can be followed individually rather than with a huge block of work. That can also be used within the "ralph wiggum" loop pattern that might help the llm to get a goal and thus iterate until goal completion. There are few articles explaining the concept if that interests you

Hope it helps a bit !