Regarding Agent Memory

Karpathy's Gist is one of those "duh" moments from the last few months everybody sort of collectively realized was reasonable ~2 months after he published it. I think once we all realized "embeddings" were just a mechanical way to match sentiment against batches of text, the magic went away and RAG was revealed to be a big waste of resources.

Some folks will say "uh, yeah, we all knew that big dawg" and the same ones have been vibing out for months accomplishing mostly the manufacture of ever increasing tool kits for their agent swarms. As far as actual utility is concerned, Karpathy's approach (just feed the model the raw data in it's entirety) is becoming the standard.

With the rise of firms like Subquadratic (SubQ?) Where massive (alleged 12 million token) contexts can be fed directly into prefill, it's hard to deny the movement of the tides. But, I'm stubborn.

There are just better ways to gain quality output with quality input orchestrated by a model, for cheap. Beads does almost all of the heavy lifting something like Jira or Obsidian would give me for ~0 effort. Holding the changelog context directly next to git just... makes sense to me. Like tying a branch to a jira card, you get change context, diffs, reasoning, and project scope tethers all right in line for agents to grep against.

There are very good reasons to not reinvent the wheel, but I think gastown did make one of the most useful context management systems available. Use it!