| name | advise |
| description | Search the required Mnemosyne knowledge repository before planning or implementation. Uses Athena's canonical dependency-resolution contract and fails if the backend cannot be prepared at ~/.agent_brain/knowledge. |
| argument-hint | <task description> |
| allowed-tools | ["Read","Bash","Grep","Glob"] |
Advise
Search durable knowledge before unfamiliar work. Mnemosyne is mandatory; never silently continue
without it.
Prepare the knowledge repository
With authenticated gh, git, and network access, prepare Mnemosyne at
$HOME/.agent_brain/knowledge by following the canonical
dependency-resolution contract exactly. Do not summarize or
override its owner precedence, trust gates, checkout rules, or revalidation requirements here.
Report the exact repository, commit SHA, and trust basis. Any resolution, authentication, checkout,
update, or revalidation failure is blocking.
Search contract
- Corpus:
skills/*.md directly under the checkout.
- Exclude
*.notes.md; use Git and pull-request history as provenance.
- Search names, descriptions, categories, tags, triggers, failed attempts, and results.
- Select at most five relevant entries, read each completely, and report its version and
verification level.
- Prefer newer, better-verified guidance and expose contradictions and failed approaches.
- Query open Mnemosyne pull requests by matched skill path and candidate title. Report an open
amendment with its branch and URL so a later Learn invocation can stack or stop instead of opening
a conflicting pull request.
Consolidation rules
- Prefer one canonical knowledge entry per user intent.
- Search Git history before recommending a name that may have been consolidated.
- Athena repository reviews use
repo-review; PR reviews use pr-review.
- Quick/default review depth is a mode, not a separate skill.
- Recommend
learn only for a new trigger, corrected command, parameter, failure mode, or verified
workflow.
Output
Return the resolved owner/Mnemosyne repository and checkout revision, then a table of entry,
version, verification, and relevance. Follow with what worked, what failed, contradictions, and
copy-ready parameters. Clearly label unverified guidance.