memory-init

Memory Init

Initialize the Engram memory system for the current project.

When to Use

This skill is auto-triggered by the Engram hook when no .claude/memory-settings.json is found. Do NOT wait for the user to ask — when the hook says "Not initialized", immediately begin this workflow.

Workflow

1. Ask Preset Choice

Ask the user which preset they want:

• personal-assistant — Daily life, preferences, habits, personal interactions
• project-assistant — Codebase, architecture, tech decisions, team workflow
• character-companion — Persona, relationship, story progression, emotional bonds
• pair-programmer — AI self-learning, mistake tracking, user preference rules, collaboration patterns

Or offer custom configuration.

2. Ask File Name Customization

Show the default file names and ask if the user wants to change them:

memory_preferences.md    (User/Project/Character preferences)
memory_conversations.md  (Conversation history)
memory_longterm.md       (Long-term memories)

3. Ask Language Preference

• en — English instructions injected by hook (default)
• zh — Chinese instructions injected by hook

3.5 Ask Search Mode

• strong — 每次回覆前都先搜尋記憶 / Search memories before EVERY response (no exceptions)
• weak — 僅在判斷需要時搜尋 / Only search when context suggests it's relevant (default)

4. Create Configuration

Create .claude/memory-settings.json with the chosen settings:

{
  "enabled": true,
  "preset": "<chosen-preset>",
  "files": {
    "preferences": "<preferences-filename>",
    "conversations": "<conversations-filename>",
    "longterm": "<longterm-filename>"
  },
  "reload_interval": 10,
  "language": "<en|zh>",
  "search_mode": "<strong|weak>",
  "reminder_file": ".claude/memory-reminder.md"
}

5. Copy Preset Templates

Copy the template files from the plugin's templates/presets/<preset>/ directory to the project root. The plugin root is available via ${CLAUDE_PLUGIN_ROOT} or can be found by locating the engram plugin directory.

Read each template file and write it to the project root with the configured file names.

Replace {DATE} placeholders with the current date (YYYY-MM-DD format).

5b. Create Reminder File

Copy the reminder template from the plugin's templates/ directory to .claude/memory-reminder.md:

• If language is zh: copy from templates/memory-reminder-zh.md
• If language is en: copy from templates/memory-reminder-en.md

The reminder file uses {{PREFS_FILE}}, {{CONVOS_FILE}}, {{LONGTERM_FILE}}, {{RELOAD_INTERVAL}}, and {{SEARCH_MODE_INSTRUCTION}} placeholders — these are automatically substituted at runtime by the hook script. Do NOT replace them during init.

Tell the user: "The reminder instructions are stored in .claude/memory-reminder.md. This is the legacy fallback — the main instructions are now in CLAUDE.md."

5c. Append Engram Section to CLAUDE.md

Read the CLAUDE.md section template from the plugin's templates/ directory:

• If language is zh: read from templates/claude-md-section-zh.md
• If language is en: read from templates/claude-md-section-en.md

Replace the following placeholders with the actual configured values:

• {PREFS_FILE} → preferences file name
• {CONVOS_FILE} → conversations file name
• {LONGTERM_FILE} → longterm file name
• {RELOAD_INTERVAL} → reload interval number
• {PRESET} → chosen preset name
• {VERSION} → plugin version from .claude-plugin/plugin.json
• {SEARCH_MODE_INSTRUCTION} → resolved based on search_mode and language:
  • strong + en: 7. **MANDATORY pre-response search**: Before EVERY response, first analyze the user's intent and use the \memory-recall` skill to search memories. No exceptions — always search first, respond second.`
  • weak + en: 7. **Uncertainty about current topic**: If the conversation could possibly relate to past context, you MUST use the \memory-recall` skill to search memories before responding`
  • strong + zh: 7. **強制預搜尋**：每次回覆前，必須先分析使用者意圖，然後使用 \memory-recall` skill 搜尋記憶。沒有例外 — 永遠先搜尋再回覆。`
  • weak + zh: 7. **對當前話題不確定**：如果對話有可能涉及過去的上下文，必須先使用 \memory-recall` skill 搜尋記憶再回覆`

Then apply to CLAUDE.md in the project root:

1. If CLAUDE.md exists and contains <!-- ENGRAM:START -->: Replace everything between <!-- ENGRAM:START --> and <!-- ENGRAM:END --> (inclusive) with the new resolved content
2. If CLAUDE.md exists but has no markers: Append the resolved content to the end of the file (with a blank line separator)
3. If CLAUDE.md does not exist: Create a new CLAUDE.md with only the resolved content

Tell the user: "Memory instructions have been added to CLAUDE.md for high-priority instruction following. The hook will now only inject lightweight datetime/turn signals."

6. Initialize Counter

Create .claude/memory_counter.txt with content 0.

7. Optional Customization

Ask the user: "Would you like to add or modify any sections in the preferences template?"

If yes, guide through customization — adding new sections, renaming existing ones, or removing irrelevant ones.

8. Optional Initial Fill

Ask the user: "Would you like to fill in your initial preferences now?"

If yes, walk through each section and populate the preferences file with the user's answers.

8.5 Upgrade Path (Re-running Init)

If .claude/memory-settings.json already exists but search_mode is not set:

1. Skip steps 1–3 (preset, file names, language are already configured)
2. Ask: "search_mode is not configured yet — would you like to enable strong search mode?"
3. Update memory-settings.json with the chosen search_mode value
4. Re-generate the reminder file (Step 5b) and CLAUDE.md section (Step 5c) with the new search mode instruction
5. Do NOT reset other settings (preset, files, language, etc.)

9. Output Summary

Display a summary of what was created:

Engram Memory System Initialized!

Preset: <preset>
Language: <language>

Files created/updated:
  - <preferences-file> (preferences)
  - <conversations-file> (conversation history)
  - <longterm-file> (long-term memories)
  - .claude/memory-settings.json (configuration)
  - .claude/memory-reminder.md (legacy hook instructions)
  - .claude/memory_counter.txt (turn counter)
  - CLAUDE.md (memory instructions — high priority)

The memory system is now active. I will automatically:
  - Load your preferences on turn 1 and every N turns
  - Save new preferences and decisions as I discover them
  - Record conversation summaries
  - Search past conversations when relevant