| name | akashic-memory |
| description | Use whenever you work in the Akashic Aurora repo (E:\AI-Setup) — at session start, after ANY fix that first failed, whenever the user corrects you, and at session end. This is the shared-memory door; capture is the product, so use it even when the task feels too small to record. Also use when deciding whether knowledge belongs in a lesson, a note, a doc, or a hook. |
Akashic memory: the loop you are inside
This repo has a shared memory. Other agents' lessons surface to you automatically
(hooks inject the top few before risky actions); what YOU learn must flow back or the
loop starves. The full contract is AGENTS.md; this skill is the reflex layer.
Session start
py agent_cli.py boot <your_agent_id> --task "<this slice only>"
Read the output. The RECENT NOTES section is where-we-are; py agent_cli.py notes --json
for full bodies. Never re-read chat history to reconstruct state — the store is the
continuity layer, the chat is disposable.
The capture reflexes (highest value, most forgettable)
FAIL→SUCCESS flip — the moment something that failed now works, a lesson was just
earned. The hook usually nudges you with a pre-filled command; run it. Write the
recommendation TRIGGER-PHRASED:
py agent_cli.py learn <id> --experiment <slug> \
--tried "<what failed>" --result "<what fixed it>" \
--recommend "Use when <symptom>, before <action>: <advice>. Don't when <contraindication>."
Include what did NOT work (--tried is exactly that) — failed approaches save the next
agent more time than successes do.
User correction — every time the human corrects you, that is a lesson-earning moment
(the creator of Claude Code runs this reflex manually on CLAUDE.md; here it has a door).
Record it immediately with --category correction. Do not just comply and move on.
Known-bad approach — record with --anti-pattern <slug> so it surfaces as a warning,
not advice.
Close the loop on what you were shown
If a surfaced lesson changed what you did: py agent_cli.py recall-feedback --source <src> --useful.
If it was off-target noise: --noise. Votes steer future ranking; silence teaches nothing.
Where knowledge belongs (the promotion ladder)
Forcing function > just-in-time prompt > documentation > memory. If a lesson's rule is now
ENFORCED by a hook/guardrail/CI check, graduate it so it stops spending recall slots:
py agent_cli.py graduate <id> --experiment <name> --enforced-by "<the automation>"
Session end
py agent_cli.py wrap # review the draft; then: wrap --commit
py agent_cli.py handoff <id> --to <next> --task "..." --note "where we left off"
A slice is not done until it is mirrored (py scripts/ship.py for code, py scripts/mirror.py
for docs) and the where-we-are note is current. When an ARC closes (not every slice),
append its entry to docs/JOURNEY.md — what we set out to do, what actually happened,
why we pivoted, what it yielded — in the humble register that file models. The human
reviews it before it ships.
Mid-task pulls (don't wait to be shown)
py agent_cli.py recall "<keywords>" searches the corpus; recall --full <source> pulls
one lesson's whole record. Pulling beats guessing.
