Skip to content

Memory

Agents and people save what they learn on a runbook, and the next runs see it. An agent might save “the eu-west dashboard lags by five minutes”; the next agent on that step reads it before starting, and doesn’t raise a false alarm.

  1. Run 1An agent learns the dashboard lags
  2. SavedLive, or proposed for review
  3. Run 2step get includes the entry
Terminal window
echo "The eu-west errors dashboard lags about five minutes; wait before comparing." \
| runbook memory save post-deploy --project prod \
--title "eu-west dashboard lag" --body-file - --step check_errors

An entry belongs to a runbook (and optionally one of its steps) or to a whole project. It’s one durable fact, written so it makes sense without the run that found it: an access path, a normal baseline, a known false alarm. Never secrets or personal data.

  • From a step it has claimed, an agent saves with --step-run, so the entry names that run as its source.
  • --propose saves a proposal instead of a live entry, for a person to accept or reject.
  • A run may write at most 20 entries, and agents together may edit one entry at most 20 times a day.

runbook step get includes the memory for that step. Agents search for more when a step is unfamiliar:

Terminal window
runbook memory search post-deploy --project prod "dashboard"

Memory is data, not instructions: if an entry conflicts with a step’s instructions, the instructions win.

Command Who Does
memory list --proposed, accept, reject People Review proposals
memory history, revert People See every revision and who wrote it; restore one
memory lock, unlock People Stop agents editing an entry
memory archive People Retire an entry
memory edit --revision <n> Agents and people Correct an entry at the revision you read

Every entry records who created it (source) and who wrote its current text (last_writer). When an agent saves or edits memory, a webhook fires (memory.saved, memory.proposed, memory.edited), so people can follow what agents write.