ditto-memory

Ditto Memory

Ditto is the user's personal long-term memory and knowledge graph (https://heyditto.ai). This skill teaches you to read and write it through the ditto CLI (the @heyditto/cli binary; heyditto is an alias). The ditto plugin also recalls relevant memories automatically on each prompt and writes salient turns back — this skill is for when you want to drive memory deliberately.

When to use

• The user says "remember…", "save this", "note that…", "for later".
• The user says "what did I…", "recall…", "have I told you…", "from my notes".
• A question is best answered from the user's prior context (preferences, decisions, people, projects) rather than general knowledge.
• The user references a topic/person/project not in this conversation but likely in their memory.

Auth — check first

Run ditto status --output json. It reports apiKey.present and source (env | config | none).

• Missing key → agent-native login (no browser/OAuth): ditto init --agent --agent-caller "Claude Code" --json. This mints a free, claimable agent account, stores the key at ~/.config/heyditto/cli/config.json, and prints a claimURL. Share the claimURL (not the API key) so the user can later own the agent. Do not ask for email/OTP/dashboard setup first.
• User has a key → ditto login <key>.
• Binary missing → npm install -g @heyditto/cli (macOS ships /usr/bin/ditto; if the wrong tool runs, use heyditto).

A read-only agent account can search/fetch but not save; write calls will be rejected — tell the user to grant write access from Settings → Agents on heyditto.ai.

Core loop: search → fetch → save

Memories are pairs (pair_id); public shares use a share_id; subjects are graph nodes (subject_id). Always pass --output json when you will parse output.

• ditto search <query>... [--include-public] [--filter-username <u>] — semantic search; multiple positional args = multiple queries (broadens recall). Returns lightweight previews + ids.
• ditto fetch <id>... [--memory-format full|outline|blocks] — full content for pair/share ids. Use outline to get block ids before a structured update.
• ditto save <content> [--source <s>] [--source-context <c>] — persist a memory. Use --source claude-code and --source-context "<repo/project>". Save proactively for durable preferences, decisions, relationships, and facts.
• ditto subjects <query> [--top-k <n>] → ditto memories <subject-id>... [--query <q>] — the cheaper way to traverse a known topic.
• ditto network <pair-id> [--limit <n>] — related memories via shared subjects.
• ditto update <id> [--content <t>|--content-file <p>|--edits-json <j>] — edit in place; prefer full-content replacement unless you have the current revision + outline block ids.
• ditto publish <id> --title "…" / ditto unpublish — only after the user explicitly asks; publish runs a privacy scan that blocks secrets by default.

Rules

• ditto status prints the live tool list straight from the server — trust it over this file if anything drifts.
• Don't fabricate memories. If search returns nothing relevant, say so.
• Don't publish or share the raw API key. Share only the claimURL.

Source + support

• CLI on npm: https://www.npmjs.com/package/@heyditto/cli
• CLI repo: https://github.com/ditto-assistant/ditto-cli
• Plugin repo: https://github.com/ditto-assistant/ditto-claude-plugins
• Support: [email protected]