Save

/save

Save knowledge from the current conversation into your Markdown knowledge base.

Memory base: ~/claude-memory — edit this line if your vault lives elsewhere. All paths below are relative to this base. Write memory files in the user's language.

Division of labor: the main model (whatever you are chatting with) does only what requires seeing the chat — extracting facts and picking the scenario. All file mechanics (folders, formats, indexes) are done by the memory-saver subagent running on a cheaper model, so long chats don't burn expensive-model usage on file I/O.

What to save — the value test

Main test for every fact: "will this line change someone's actions two weeks from now?" If not — don't save it.

Priority (top first):

1. Invariants and rules — "always X, never Y, because Z". The most valuable kind: it prevents future mistakes
2. Root causes, not symptoms — what exactly was broken and why the fix is correct. "Fixed the sync bug" is noise; "pull compared raw files against redacted ones, hence eternal conflicts" is knowledge
3. Decisions with rejected alternatives — "X instead of Y, because Z", so you never re-evaluate Y from scratch
4. Exact identifiers — commands, paths, IDs, ports, versions, branch/PR names: things impossible to recall from memory
5. Open loops 🔴 — undeployed fixes, unverified assumptions, contradictions
6. Tools, links, vocabulary worth keeping

Do NOT save:

• The journey ("tried A, then B, then C") — only the outcome and the root cause; exception: a dead end that is tempting to retry (then: why it fails)
• Anything obvious from the code or git history
• General knowledge about technologies — the model already knows it
• Chit-chat, "ok/got it", intermediate versions of decisions, duplicates of already-saved facts

Check before writing: every fact must contain something concrete — a number, a path, a name, a reason. A fact without specifics is a paraphrase; drop it or sharpen it.

Secrets

NEVER write API keys, tokens, passwords or connection strings into memory — replace them with <<REDACTED>>. The memory base may live in a git repository and sync between machines.

Algorithm (main model)

1. Find the boundary

Look for a previous /save invocation in this chat's history.

• Found → save ONLY what happened after the last one
• Not found → save everything important from the whole conversation

⚠️ In a very long chat, early details may already be compacted away by context summarization. If you can see the start of the conversation was trimmed — say so honestly in your report and suggest saving mid-session next time.

2. Pick the scenario (only chat history can tell)

• A — first /save in this chat → new file
• B — repeat /save in the same chat → append to the file linked in your previous save report
• C — new chat, new topic → new independent file. Do NOT attach to existing files unless the user explicitly asks
• D — user explicitly said "this continues X" → part-file (chain)

2.5 Identify the project and folder (BEFORE the brief — only you see the chat)

The rule is "one project = one folder" — details in topics.md. If the conversation is about a specific project:

• the project already has a folder → pass it in the brief
• the project is NEW (its own repo/bot/deliverable, even on a familiar stack) → ask the user: Create a new folder <Name/> for this project? and pass the answer in the brief
• in doubt whether it's part of [X] or separate → ask. Never merge two projects silently

Ask the user BEFORE calling the subagent — it cannot see the chat and cannot ask follow-ups.

3. Build the brief and call the subagent

Call the Agent tool: subagent_type: "memory-saver", run_in_background: false. Put into the brief EVERYTHING the subagent cannot learn on its own (it does not see the chat):

Date: YYYY-MM-DD (today's real date)
Scenario: A/B/C/D. For B/D — target file: [[Folder/name]]
Folder: <exact folder from step 2.5; for a new project — "create new folder <Name/>">
Topic: one line
Facts to save:
- [filtered by the priorities above — specifics: commands, paths, decisions, root causes]
Open problems 🔴: [undeployed / unverified / contradictions — or "none"]
Multiple topics → list them separately: the subagent creates one file per topic.

Write the facts out in full — the subagent records ONLY what is in the brief and invents nothing.

4. Relay the subagent's report

One-two lines from its result: ✅ Created [[...]] / ✅ Appended /SAVE #N to [[...]]. If the subagent returned a question (e.g. no matching folder) — ask the user and call it again.

Fallback — if the Agent tool is unavailable

Do everything yourself using the full rules: algorithm.md (scenarios, part-files), examples.md (formats), topics.md (folder mapping). Key rules: append-only; N = count("^## /SAVE #") + 2; only updated: changes in frontmatter; wikilinks in frontmatter are quoted; main.md — one file = one row ≤ 200 chars, rows only inside the table; Days files are # YYYY-MM-DD with no frontmatter; 📌 block when a file has 3+ save sections or an open problem 🔴.

Conventions

• Files: lowercase kebab-case, English filenames
• Wikilinks: [[Folder/name]] with the full path, no .md
• Dates: ISO YYYY-MM-DD
• File contents: the user's language