a second brain, week two
01 May 2026a week ago i wrote about a memory MCP i was using as a second brain. a few notes from week two.
most of my actual work this week went into an evaluation framework i’m building, and the memory system was on for all of it (and for the rest of what was going on). three things i noticed.
there’s a low-grade dread when i’m working: that no matter how diligent i am about notes, some lessons or threads will get lost. not because i didn’t write them down, but because i didn’t have a system that would surface them at the right time. this week that pressure was noticeably lighter. lessons get captured, recalled when relevant, and (maybe the most valuable part) they travel between different work surfaces. the memory acts as connective tissue (extracellular matrix is the picture in my head), so i can be largely agnostic to which agentic system i’m in: claude code, codex, the interview skill, claude desktop. it’s the second brain thing, more or less. it freed up bandwidth in a way i didn’t expect, which is itself a signal that whatever i had in place before wasn’t doing as much work as i thought it was.
another shift: mental reviews are cheap now. what previously took a deliberate sit-down: gather context, find the notes, get into the right headspace, is now one question in any chat. that’s a pro and a con. lower barrier means i actually do it, more often. but each review is also shallower than the old kind, and it surfaces only what got captured, so the meaningful undocumented stuff (a conversation that shifted how i think about a problem) stays invisible. net positive, since the alternative was skipping reviews entirely (but worth naming).
the last thing is where i’m pushing back, gently, on tim kellogg’s suggestion: store only the fact that claude went wrong, and let an ambient process fix the system. i get the appeal, but i’m not there. a lot of what gets captured as “claude failures” are things i want myself to be better at, not just the agent. if the system auto-patches around them, i lose the signal that would have made me learn from them. control over what gets captured, what counts as wrong, and how it gets fixed is part of the value (at least for now), not friction to remove (that may shift as i trust the system more).
(this is the upbeat version, partly because it’s a sunny friday morning. the friction-makes-learning question from last week is still open and i don’t want to pretend otherwise.)
P.S. two small workflow changes this week:
- added a grill-me move (à la matt pocock) to my interview skill, still using it constantly
- added an end-of-session summary + warm-start (with a hemingway bridge) to my CLAUDE.md
P.P.S. tags are still something i’m not getting right: here on sparse thoughts, and in the memory system. open problem.