Memory Commit

Memory Commit

Manual capture of a single decision, preference, fact, skill, task, or conversation snippet into mnemo. Enforces the "WHY not just WHAT" rule and chooses the correct context_type so retrieval is precise.

When to Use

Trigger on explicit user signals:

• "remember this", "save this", "save for next time"
• "ghi nho", "luu lai", "nho lai"
• "let's commit this to memory", "store this"
• After a deliberate decision: "we picked X because Y"
• After a stated preference: "I always want X"
• After correcting the agent: high-value capture (prevents repeat)

Do NOT trigger on incidental mentions or speculative options not yet chosen.

Steps

1. Identify the content the user wants to remember:

   • Preceding 1-3 messages (default), OR
   • A specific quoted span the user references, OR
   • The current selection if invoked via slash command

2. Determine context_type via this decision tree:

   Signal	context_type
   "we decided", "we picked X over Y", "going with"	decision
   "I prefer", "I always want", "default to"	preference
   "X is at version Y", env vars, API shapes, file locations	fact
   "to do X you run Y then Z", procedure, how-to	skill
   "todo", "remember to", deadline, "by Friday"	task
   none of the above (general context)	conversation

   When ambiguous, ask the user one short question. Do not silently default to conversation for high-signal content.

3. Compose the capture text with WHY included:

   • BAD: "Use Polars"
   • GOOD: "Use Polars for dataframes (not pandas) because the codebase processes 50M-row datasets and Polars is 20x faster than pandas in our benchmark."

4. Capture via the typed action:

   memory(action="capture",
          text="<composed text>",
          context_type="<chosen type>",
          category="<project-name or topic>",
          tags=["<topic>", "<scope>"])
   

5. Confirm to the user:

   Saved as <context_type>. Memory ID: <id>.
   (Deduplicated against existing similar memory: <existing_id>.)  # if applicable
   

   If the response includes deduplicated: true, surface the existing memory ID so the user can decide whether to update it instead.

6. Verify retrieval for high-stakes captures (decisions, corrections):

   memory(action="search", query="<the keywords a future agent would use>", limit=3)
   

   If the just-captured memory does not appear in top-3 results, rewrite with better keywords and re-capture (delete the original first).

Quality Rules

• WHY not WHAT: every captured memory must explain reasoning.
• One insight per capture: do not cram multiple unrelated facts into one entry - they will not retrieve well.
• Specific over generic: include version numbers, file paths, percentages where relevant.
• No ephemeral state: do not save "currently on line 42 of file X" or "the test is failing right now" - these will be wrong tomorrow.
• Honor user words: when the user dictates exact phrasing for a preference or decision, use their words verbatim.

Examples

User: "Remember that we always use ruff format, not black."

memory(action="capture",
       text="Use ruff format (not black) for Python formatting because ruff is the project standard - faster, single-binary, and configured in pyproject.toml.",
       context_type="preference",
       category="tooling",
       tags=["python", "formatting"])

User: "Ghi nho: deploy len VM dung make up-, khong dung docker compose truc tiep."

memory(action="capture",
       text="VM deployment: dung make up-<service> / make down-<service> de inject secrets tu skret SSM qua RUN macro. KHONG chay docker compose truc tiep tren VM (mat secret injection).",
       context_type="skill",
       category="deployment",
       tags=["vm", "make", "skret"])