So how are people handling this in real setups?

do you delete old AI-generated notes, mark them stale, keep source-only archives, use review queues, or just let the context pile grow?

For people building agents: when something goes wrong, do you record it as:

- eval case
- prompt/rule change
- memory
- tool/hook
- human process
- “ignore, one-off”

where does it do the processing? and how do i structure the workspace for it?

had many good conversations afterwards, on a roll and i hope to keep the energy going. thanks @zachkrall.com for the invite and @modal-labs.bsky.social for hosting!

handing out these ai receipts as mini uxr sessions taught me a lot about how to design ai memory to draw connections that feel hospitable. It's enzyme's unfair advantage.

i contrasted preference memory -- helping you accomplish a task you asked for, from conviction memory -- which wonders with you about what you care to find out

with a challenge: you'd forgive AI slop on a charming receipt, but not when it comes to your daily AI tools

i shared learnings that started with a months long experiment: seeing how people reacted to sequences of my personal reading highlights printed on thermal receipts, in exchange for themes in their uploaded photos.

3.1k lines of rust. 5MB binary. records both sides of the call. no accounts, no cloud, your audio never leaves your machine. apache 2.0

brew install jshph/aside/aside

github.com/jshph/aside/

ran it on two meetings today. the note it wrote linked back to a half-finished doc i forgot about — not as action items, as context for what i'd been thinking about

i was tired of pasting transcripts into claude myself. so i built the integration layer

the distillation is a 240-line markdown file — a claude code skill the LLM follows as instructions. your memo lines are the attention signal, your vault is the context window

you can read it, fork it, change how it writes notes. no black box