Agent skill

recall-write

Write a note to memory with write gate evaluation. Only saves behavior-changing facts.

Stars 192
Forks 12

Install this agent skill to your Project

npx add-skill https://github.com/davegoldblatt/total-recall/tree/main/skills/recall-write

SKILL.md

Write a note to memory, applying the write gate. Default destination: daily log.

The user's note: $ARGUMENTS

Write Gate Protocol

Before writing, evaluate:

  1. Does it change future behavior? (preference, boundary, recurring pattern) → WRITE
  2. Is it a commitment with consequences? (deadline, deliverable, follow-up) → WRITE
  3. Is it a decision with rationale? (why X over Y, worth preserving) → WRITE
  4. Is it a stable fact that will matter again? (not transient, not obvious) → WRITE
  5. Did the user explicitly say "remember this"? → ALWAYS WRITE

If NONE of these are true, tell the user why it didn't pass the gate. Suggest /recall-log for raw capture without the gate.

Default: Write to Daily Log

ALL writes go to memory/daily/YYYY-MM-DD.md first. Create the file if it doesn't exist:

markdown
# YYYY-MM-DD

## Decisions

## Corrections

## Commitments

## Open Loops

## Notes

Append a timestamped entry under the appropriate section:

[HH:MM] note text here

Suggest Promotion (Don't Auto-Promote)

After writing to the daily log, if the note seems durable, suggest where it could be promoted — but don't do it automatically:

Written to memory/daily/2026-02-05.md:
  [14:32] User prefers bullet points over prose for summaries

This looks like a lasting preference. Want me to also promote it to:
  → memory/registers/preferences.md
  → CLAUDE.local.md (if it should affect every session)

The user decides. If they say yes, write to the register with metadata:

markdown
- **claim**: [the fact/preference/decision]
- **confidence**: high | medium | low
- **evidence**: [how we know — user said, observed, corrected]
- **last_verified**: [today's date]

Exceptions: Direct Promotion

These can skip the "suggest" step and promote directly (but still write to daily log too):

  1. Explicit corrections — user corrects a prior memory entry. Update the register immediately, mark old entry as [superseded: date].
  2. Explicit "remember this" — user clearly wants it stored durably. Write to the appropriate register.
  3. Deadline/commitment — always goes to registers/open-loops.md AND daily log.

Contradiction Check

Before writing, quickly check existing memory for related claims:

  1. Check CLAUDE.local.md
  2. Check relevant register (if promoting)

If a contradiction is found:

  • Show the user: "Existing memory says [X]. You're now saying [Y]. Update?"
  • Mark old entry as [superseded: date] with reason
  • Write new entry

ID Assignment on Promotion

When writing an entry to CLAUDE.local.md or a register (either via direct promotion or user-confirmed promotion), assign a durable ID:

  1. Generate an ID: ^tr + 10 random lowercase hex characters
  2. Check for collisions against existing IDs in memory/.recall/metadata.json (if it exists) and any inline IDs visible in the destination file
  3. Append ^[id] to the end of the entry line (for single-line list items starting with - )
  4. Create or update memory/.recall/metadata.json with an entry for the new ID:
    json
    {
      "created_at": "[ISO 8601 timestamp]",
      "last_reviewed_at": "[ISO 8601 timestamp]",
      "pinned": false,
      "snoozed_until": null,
      "status": "active",
      "tier": "working"
    }
    
    Set tier to "working" for CLAUDE.local.md or "register" for register files.
  5. Create the memory/.recall/ directory if it doesn't exist
  6. Write metadata.json with sorted keys and 2-space indentation

Daily log entries do NOT get IDs. Only entries promoted to CLAUDE.local.md or registers are tagged.

Multi-line metadata blocks (claim/confidence/evidence/last_verified format) are not tagged with IDs in v1. Only single-line list items get IDs.

After Writing

Confirm to the user:

Noted in memory/daily/[date].md: [one-line summary]

If also promoted: mention the additional destination and the assigned ID.

Expand your agent's capabilities with these related and highly-rated skills.

Didn't find tool you were looking for?

Be as detailed as possible for better results