| name | memory-update |
| description | Scan agent's session-history transcripts for save-worthy signals (corrections, preferences, decisions, references), propose and write auto-memory files with valid frontmatter and MEMORY.md entry. Use when the user says "save this to memory", "remember that", or "scan this session for memories". |
Scan session history, surface save-worthy moments, propose → confirm → write. Never fabricate — every proposal must cite a transcript turn.
Scope
Creates or revises memory files in the agent's memory directory. Does not delete, merge-stale, or audit existing files (that is memory-clean). Does not redact or sanitize (that is memory-sanitize).
Path resolution
SKILL_SCRIPTS="${MEMORY_UPDATE_SKILL_SCRIPTS:-$HOME/.claude/claude/skills/memory-update/scripts}"
MEMORY_DIR=$("$SKILL_SCRIPTS/resolve-paths.sh" memory_dir)
SESSION_HISTORY_GLOB=$("$SKILL_SCRIPTS/resolve-paths.sh" session_history_glob)
Abort on non-zero exit. Override MEMORY_UPDATE_SKILL_SCRIPTS if installed outside $HOME/.claude. Override MEMORY_DIR / SESSION_HISTORY_GLOB env vars to target a non-Claude-Code agent's layout.
Memory types and required sections
Read references/MEMORY-FRONTMATTER.md for the full schema. Key rules:
| type | Required extra sections |
|---|
feedback | **Why:** and **How to apply:** |
project | **Why:** and **How to apply:** |
user | none beyond frontmatter |
reference | none beyond frontmatter |
Filename: <type>_<slug>.md. Slug: lowercase, hyphens only, ≤ 40 chars.
Workflow
1. Resolve paths (above)
2. Scan session history for proposals
Run scripts/scan-session.sh "$SESSION_HISTORY_GLOB" → JSON array:
[{ "type": "feedback", "slug": "prefer-foo", "evidence_turn_ids": ["uuid-123"],
"draft_body": "...", "draft_index_entry": "- [Title](file.md) — hook" }, ...]
Read references/SIGNAL-HEURISTICS.md to understand which turns qualify. If the user triggered the skill with an explicit instruction ("remember that X"), synthesize a proposal from that instruction directly instead of scanning.
3. Present proposals and wait for confirmation
Render each proposal as:
Proposal N — type: feedback
Evidence: turn uuid-123 ("the user said 'stop doing X'")
Draft:
---
name: …
description: …
type: feedback
---
Rule text. **Why:** … **How to apply:** …
Index entry: - [Title](file.md) — hook (N chars)
Accept / Reject / Edit?
Wait for explicit user response per proposal. Never write without confirmation.
4. Check for near-duplicates before writing
rg -l -F "<key phrase from draft>" "$MEMORY_DIR"
rg works on any directory regardless of git worktree boundaries. If a hit exists, show the existing file's content and ask whether to merge into it or create new.
5. Write
- Write
"$MEMORY_DIR/<type>_<slug>.md" with frontmatter + body.
- Append (or update) one-line entry in
"$MEMORY_DIR/MEMORY.md" (≤ 150 chars).
6. Verify
Read back each written file; confirm frontmatter parses as YAML (key: value lines present, no tab-indentation errors). Report which files were written and their line counts.
Anti-patterns (never do)
- Fabricate a memory without citing a transcript turn or explicit user statement.
- Write without per-proposal user confirmation.
- Store code patterns, fix recipes, git history, or ephemeral task state.
- Create MEMORY.md entries exceeding 150 chars.
