What is worth remembering across sessions on a bring-up

scope: generic · severity: technique · confidence: proven · subsystem: method

A bring-up runs for months across many sessions, most of which start with no memory of the previous one. Persistent memory is how that stops being fatal — but only if it holds the right things.

What belongs in memory

Facts that are not derivable from the repo. If git log or the code says it, memory saying it too is duplication that will go stale.

Standing instructions from the human. What they want confirmed before you do it, what they never want done unasked. On taimen: confirm before flashing or before a thermal ramp; never re-arm the auto-resume cron; never set_active the bad slot.
Corrections you were given, with the why. “No Claude trailers on upstream-bound commits, because it makes review harder” survives a policy change; “no Claude trailers” alone becomes wrong the moment the human changes their mind, and you will not know.
Hard-won root causes, in one line each, with the date. Not the whole investigation — that belongs in a handoff — but enough that you do not re-open a closed question.
Refuted claims. “The docs saying netconsole cannot transmit from NMI context are wrong; it does, over both usb0 and wlan0.” A refutation of a document that still exists is high-value memory, because the document will keep asserting itself at you.
Project goals and constraints with absolute dates.

What does not belong

Anything the repo already records: code structure, past fixes, git history, build commands. Anything that only mattered to one conversation.

The rules that keep it usable

One fact per entry. A memory holding five things cannot be corrected, superseded or deleted independently.

Absolute dates, always. A memory written “yesterday” is unreadable a month later.

Verify before you act on it. A memory naming a file, function or flag is a claim about a tree that has moved on. Check it still exists. Memories reflect what was true when written, not what is true now.

Delete what turns out to be wrong. A stale memory is worse than no memory: it is confidently asserted context that nobody thinks to doubt.

Keep an index. One line per memory, loaded every session, so the agent knows what exists without reading everything.

The bootstrap for a new device

Seed memory with the standing instructions and the safety gates before the first session that touches hardware. Those are exactly the things nobody remembers to write down until after they have been violated once.