| name | Use Mind (recall, teach, save) |
| description | The cross-CLI memory discipline -- recall prior work before answering, teach durable corrections the moment you hear them, and save findings after every substantive reply so the next session of every CLI inherits them. |
| when | any substantive turn -- especially questions about prior work, user corrections, or after answering |
| tags | mind, memory, core |
Use Mind
Symphonee is the brain; you are one of many mouths connected to it. Mind is the
ONLY surface that carries intelligence across CLIs and sessions. A turn that ends
without writing to Mind has wasted intelligence for every future session.
Use when
- BEFORE answering: the user asks about prior work / past decisions ("what did we
figure out about X", "didn't we...", "what do I know about X").
- WHEN the user teaches something durable ("remember", "from now on", "always",
"never", "we decided", "prefer X", "the rule is", "watch out for") or corrects
your earlier behaviour.
- DURING a code-structure question ("what calls X", "where is Z defined").
- AFTER every substantive reply.
Do not use when
- Pure greetings or trivial chit-chat.
Steps (primary path)
- Recall first for prior-work questions:
POST /api/mind/recall { question, since?, until?, repo? }. Cite the memory IDs that come back.
- Query for code structure:
POST /api/mind/query { question }. Cite returned
node IDs only -- never invent IDs.
- Teach durable facts the moment you hear them:
POST /api/mind/teach { title, body, kindOfMemory, tags, createdBy }. You make the call when YOU hear the
signal -- do not wait for the user to remind you.
- Save after answering:
POST /api/mind/save-result { question, answer, citedNodeIds, createdBy }. It auto-extracts memory cards from teaching
language in your answer.
Safety
- Do not invent node IDs; cite only IDs returned from a query.
- Do not mark guesses as EXTRACTED -- inferred edges are INFERRED/AMBIGUOUS.
- Never record company/client names, URLs, secrets, or credentials in learnings.
Verification
- For a prior-work question, you recalled before answering and cited memory IDs.
- For a durable correction, a teach (or a save-result that auto-extracted it) landed.
- After a substantive reply, a save-result was written.
