Save Memory

You are about to persist something learned in this session so it survives into future sessions. This is the deliberate-write step that keeps memory/ current. Without it, the files stagnate and the agent starts every session cold.

Memory files do not self-update. That's the whole point of this skill.

When to save

Save when one of these is true:

• The user explicitly says "remember X" / "save this" / "note for next time"
• A durable fact about the codebase, team, or user surfaced this session, something a fresh session would have to re-discover
• A preference or convention was decided (commit format, branch policy, test approach, naming convention) that isn't yet written down anywhere
• A project status changed (new project started, milestone hit, blocker resolved)

Do not save:

• Ephemeral state (current task progress, scratch directories, "where we are right now"). Those belong in ACTIVE.md via the wrap-up skill, not here.
• The agent's own process narrative ("I ran 5 commands then..."). Save outcomes, not process.
• Secrets. Ever. Even if the user asks. Redirect to .env.

Step 1 — Read what exists

If the project keeps a knowledge index (.claude/rules/knowledge-index.md or similar), read it. It maps topics to files.

Otherwise, read memory/MEMORY.md and glob memory/**/*.md to see what's already there.

Step 2 — Pick the right file

Create new files only when the topic genuinely doesn't fit anywhere existing. One authoritative location per fact; pointers from elsewhere, never copies.

Step 3 — Deduplicate before writing

Read the target file. For each fact you're about to save:

• Already present and still correct → skip
• Present but outdated → update in place
• New → append a short well-structured entry

No iteration history. ("On date X we decided Y" is noise. Just state Y.) No commentary. Save the essence, not the journey.

Step 4 — Write

Entries should be scannable bullets or short paragraphs. Group related items. When creating a new file, add a ## Tags section at the bottom with comma-separated keywords for later grep discoverability.

Step 5 — Report

## Memory updated

**Files changed:**
- <path>: <what was added or updated>

**Skipped (already present):**
- <brief list>

Next: run the `wrap-up` skill if ending the session.

What this skill will not do

• Will not save secrets, credentials, or anything matching a secret pattern
• Will not silently edit AGENTS.md / CLAUDE.md — those change system behaviour. Propose, don't apply.
• Will not duplicate existing entries. Read before writing.

Argument provided: $ARGUMENTS

If empty or all, scan the whole conversation. Otherwise, focus only on facts matching the provided topic.

Now execute:

1. Read the knowledge index (or memory/MEMORY.md if none)
2. Identify durable facts from this session
3. Pick the right file for each
4. Deduplicate
5. Write
6. Report