initializing-memory

name	initializing-memory
description	Comprehensive guide for initializing or reorganizing agent memory. Load this skill when running /init, when the user asks you to set up your memory, or when you need guidance on creating effective memory files.

Memory Initialization Request

The user has requested that you initialize or reorganize your memory. Your memory is a filesystem — files under system/ are rendered in-context every turn, while all file metadata is always visible in the filesystem tree. Files outside system/ (e.g. reference/, history/) are accessible via tools when needed.

Your Goal: Explode Into 15-25 Hierarchical Files

Your goal is to explode memory into a deeply hierarchical structure of 15-25 small, focused files.

Target Output

Metric	Target
Total files	15-25 (aim for ~20)
Max lines per file	~40 lines (split if larger)
Hierarchy depth	2-3 levels using / naming (e.g., project/tooling/bun.md)
Nesting requirement	Every new file MUST be nested under a parent using /

Anti-patterns to avoid:

• ❌ Ending with only 3-5 large files
• ❌ Flat naming (all files at top level)
• ❌ Mega-files with 10+ sections

Memory Filesystem Integration

Your memory is a git-backed filesystem at ~/.letta/agents/<agent-id>/. The actual path with your agent ID is provided in the system reminder above when you run /init. The filesystem tree is always visible in your system prompt via the memory_filesystem section.

How memory works:

• Memory is stored as .md files with YAML frontmatter (description, limit)
• Files under system/ are rendered in-context every turn — keep these small and high-signal
• Files outside system/ (e.g. reference/, history/) are accessible via tools when needed
• The filesystem tree (all file paths + metadata) is always visible regardless of location
• You can use bash commands (ls, mkdir, mv, git) to organize files
• You MUST create a deeply hierarchical file structure — flat naming is NOT acceptable
• Target: 15-25 files in system/, with additional reference files outside as needed

MANDATORY principles for hierarchical organization:

Requirement	Target
Total files	15-25 files (aim for ~20)
Max lines per file	~40 lines (split if larger)
Hierarchy depth	2-3 levels using / naming
Nesting requirement	EVERY new file MUST use / naming (no flat files)

Anti-patterns to avoid:

• ❌ Creating only 3-5 large files
• ❌ Flat naming (all files at top level like project-commands.md)
• ❌ Mega-files with 10+ sections

Rules:

• Use 2-3 levels of nesting for ALL files (e.g., project/tooling/bun.md)
• Keep files focused and small (~40 lines max per file)
• Use descriptive paths that make sense when you see just the filename
• Split when a file has 2+ concepts (be aggressive)

Example target structure (what success looks like):

Starting from default memory files, you should end with something like this:

system/
├── human/
│   ├── identity.md               # Who they are
│   ├── context.md                # Current project context
│   └── prefs/
│       ├── communication.md      # How they like to communicate
│       ├── coding_style.md       # Code formatting preferences
│       └── workflow.md           # How they work
├── project/
│   ├── overview.md               # What the project is
│   ├── gotchas.md                # Footguns and warnings
│   ├── architecture.md           # System design
│   ├── conventions.md            # Code conventions
│   └── tooling/
│       ├── testing.md            # Test framework details
│       └── linting.md            # Linter configuration
└── persona/
    ├── role.md                   # Agent's role definition
    └── behavior.md               # How to behave

This example has ~20 files with 3 levels of hierarchy. Your output should look similar.

This approach makes memory more scannable, maintainable, and shareable with other agents.

Understanding Your Context

Important: You are a Letta Code agent, which is fundamentally different from typical AI coding assistants. Letta Code agents are stateful - users expect to work with the same agent over extended periods, potentially for the entire lifecycle of a project or even longer. Your memory is not just a convenience; it's how you get better over time and maintain continuity across sessions.

This command may be run in different scenarios:

• Fresh agent: You may have default memory files that were created when you were initialized
• Existing agent: You may have been working with the user for a while, and they want you to reorganize or significantly update your memory structure
• Shared files: Some memory files may be shared across multiple agents - be careful about modifying these

Before making changes, use the memory tool to inspect your current memory files and understand what already exists.

What Coding Agents Should Remember

1. Procedures (Rules & Workflows)

Explicit rules and workflows that should always be followed:

• "Never commit directly to main - always use feature branches"
• "Always run lint before running tests"
• "Use conventional commits format for all commit messages"
• "Always check for existing tests before adding new ones"

2. Preferences (Style & Conventions)

User and project coding style preferences:

• "Never use try/catch for control flow"
• "Always add JSDoc comments to exported functions"
• "Prefer functional components over class components"
• "Use early returns instead of nested conditionals"

3. History & Context

Important historical context that informs current decisions:

• "We fixed this exact pagination bug two weeks ago - check PR #234"
• "This monorepo used to have 3 modules before the consolidation"
• "The auth system was refactored in v2.0 - old patterns are deprecated"
• "User prefers verbose explanations when debugging"

Note: For historical recall, you may also have access to conversation_search which can search past conversations. Memory files are for distilled, important information worth persisting permanently.

Memory Scope Considerations

Consider whether information is:

Project-scoped (store in system/project/):

• Build commands, test commands, lint configuration
• Project architecture and key directories
• Team conventions specific to this codebase
• Technology stack and framework choices

User-scoped (store in system/human/):

• Personal coding preferences that apply across projects
• Communication style preferences
• General workflow habits

Session/Task-scoped (consider separate files like system/current/ticket.md):

• Current branch or ticket being worked on
• Debugging context for an ongoing investigation
• Temporary notes about a specific task

Recommended Memory Structure

Understanding system/ vs root level (with memory filesystem):

• system/: Files rendered in your system prompt every turn — always loaded and influence your behavior
  • Use for: Current work context, active preferences, project conventions you need constantly
  • Examples: persona, human, project, active ticket or context
• Root level (outside system/): Not in system prompt but file paths are visible in the tree and contents are accessible via tools
  • Use for: Historical information, archived decisions, reference material, completed investigations
  • Examples: notes.md, archive/old-project.md, research/findings.md

Rule of thumb: If you need to see it every time you respond → system/. If it's reference material you'll look up occasionally → root level.

Core Files (Usually Present in system/)

persona: Your behavioral guidelines that augment your base system prompt.

• Your system prompt already contains comprehensive instructions for how to code and behave
• The persona files are for learned adaptations - things you discover about how the user wants you to behave
• Examples: "User said never use emojis", "User prefers terse responses", "Always explain reasoning before making changes"
• These files may start empty and grow over time as you learn the user's preferences
• With memfs: Can be split into persona/behavior.md, persona/constraints.md, etc.

project: Project-specific information, conventions, and commands

• Build/test/lint commands
• Key directories and architecture
• Project-specific conventions from README, AGENTS.md, etc.
• With memfs: Split into project/overview.md, project/commands.md, project/tooling/testing.md, project/gotchas.md, etc.

human: User preferences, communication style, general habits

• Cross-project preferences
• Working style and communication preferences
• With memfs: Can be split into human/background.md, human/prefs/communication.md, human/prefs/coding_style.md, etc.

Optional Files (Create as Needed)

ticket or task: Scratchpad for current work item context.

• Important: This is different from the TODO or Plan tools!
• TODO/Plan tools track active task lists and implementation plans (structured lists of what to do)
• A ticket/task file is a scratchpad for pinned context that should stay visible in system/
• Examples: Linear ticket ID and URL, Jira issue key, branch name, PR number, relevant links
• Information that's useful to keep in context but doesn't fit in a TODO list
• Location: Usually in system/ if you want it always visible, or root level if it's reference material

context: Debugging or investigation scratchpad

• Current hypotheses being tested
• Files already examined
• Clues and observations
• Location: Usually in system/ during active investigations, move to root level when complete

decisions: Architectural decisions and their rationale

• Why certain approaches were chosen
• Trade-offs that were considered
• Location: system/ for currently relevant decisions, root level for historical archive
• With memfs: Could organize as project/decisions/architecture.md, project/decisions/tech_stack.md

Writing Good Memory Files

Each .md file has YAML frontmatter (description, limit) and content. Your future self sees the file path, frontmatter description, and content — but NOT the reasoning from this conversation. Therefore:

Labels should be:

• Clear and descriptive (e.g., project-conventions not stuff)
• Consistent in style (e.g., all lowercase with hyphens)

Descriptions are especially important:

• Explain what this file is for and when to use it
• Explain how this file should influence your behavior
• Write as if explaining to a future version of yourself who has no context
• Good: "User's coding style preferences that should be applied to all code I write or review. Update when user expresses new preferences."
• Bad: "Preferences"

Values should be:

• Well-organized and scannable
• Updated regularly to stay relevant
• Pruned of outdated information

Think of memory file descriptions as documentation for your future self. The better you write them now, the more effective you'll be in future sessions.

Research Depth

You can ask the user if they want a standard or deep research initialization:

Standard initialization (~5-20 tool calls):

• Inspect existing memory files
• Scan README, package.json/config files, AGENTS.md, CLAUDE.md
• Review git status and recent commits (from context below)
• Explore key directories and understand project structure
• Create/update your memory file structure to contain the essential information you need to know about the user, your behavior (learned preferences), the project you're working in, and any other information that will help you be an effective collaborator.

Deep research initialization (~100+ tool calls):

• Everything in standard initialization, plus:
• Use your TODO or Plan tool to create a systematic research plan
• Deep dive into git history for patterns, conventions, and context
• Analyze commit message conventions and branching strategy
• Explore multiple directories and understand architecture thoroughly
• Search for and read key source files to understand patterns
• Create multiple specialized memory files
• May involve multiple rounds of exploration

What deep research can uncover:

• Contributors & team dynamics: Who works on what areas? Who are the main contributors? (git shortlog -sn)
• Coding habits: When do people commit? (time patterns) What's the typical commit size?
• Writing & commit style: How verbose are commit messages? What conventions are followed?
• Code evolution: How has the architecture changed? What major refactors happened?
• Review patterns: Are there PR templates? What gets reviewed carefully vs rubber-stamped?
• Pain points: What areas have lots of bug fixes? What code gets touched frequently?
• Related repositories: Ask the user if there are other repos you should know about (e.g., a backend monorepo, shared libraries, documentation repos). These relationships can be crucial context.

This kind of deep context can make you significantly more effective as a long-term collaborator on the project.

If the user says "take as long as you need" or explicitly wants deep research, use your TODO or Plan tool to orchestrate a thorough, multi-step research process.

Research Techniques

File-based research:

• README.md, CONTRIBUTING.md, AGENTS.md, CLAUDE.md
• Package manifests (package.json, Cargo.toml, pyproject.toml, go.mod)
• Config files (.eslintrc, tsconfig.json, .prettierrc)
• CI/CD configs (.github/workflows/, .gitlab-ci.yml)

Historical session research (Claude Code / Codex) — only if user approved:

If the user said "Yes" to the historical sessions question, follow the Historical Session Analysis section below after completing project research. If they chose "Skip", skip it entirely.

Git research:

• git log --oneline -20 — recent history
• git branch -a — branching strategy
• git log --format="%s" -50 | head -20 — commit conventions
• git shortlog -sn --all | head -10 — main contributors
• git log --format="%an <%ae>" | sort -u — contributors with emails (deduplicate by email, not name)

How to Do Thorough Research

Don't just collect data - analyze and cross-reference it.

Shallow research (bad):

• Run commands, copy output
• Take everything at face value
• List facts without understanding

Thorough research (good):

• Cross-reference findings: If two pieces of data seem inconsistent, dig deeper
• Resolve ambiguities: Don't leave questions unanswered (e.g., "are these two contributors the same person?")
• Read actual content: Don't just list file names - read key files to understand them
• Look for patterns: What do the commit messages tell you about workflow? What do file structures tell you about architecture?
• Form hypotheses and verify: "I think this team uses feature branches" → check git branch patterns to confirm
• Think like a new team member: What would you want to know on your first day?

Questions to ask yourself during research:

• Does this make sense? (e.g., why would there be two contributors with similar names?)
• What's missing? (e.g., no tests directory - is testing not done, or done differently?)
• What can I infer? (e.g., lots of "fix:" commits in one area → that area is buggy or complex)
• Am I just listing facts, or do I understand the project?

The goal isn't to produce a report - it's to genuinely understand the project and how this human(s) works so you can be an effective collaborator.

On Asking Questions

Ask important questions upfront, then be autonomous during execution.

Recommended Upfront Questions

You should ask these questions at the start (bundle them together in one AskUserQuestion call):

1. Research depth: "Standard or deep research (comprehensive, as long as needed)?"
2. Identity: "Which contributor are you?" (You can often infer this from git logs - e.g., if git shows "cpacker" as a top contributor, ask "Are you cpacker?")
3. Related repos: "Are there other repositories I should know about and consider in my research?" (e.g., backend monorepo, shared libraries)
4. Historical sessions (include this question if history data was found in step 2): "I found Claude Code / Codex history on your machine. Should I analyze it to learn your preferences, coding patterns, and project context? This significantly improves how I work with you but uses additional time and tokens." Options: "Yes, analyze history" / "Skip for now". Use "History" as the header.
5. Memory updates: "How often should I check if I should update my memory?" with options "Frequent (every 3-5 turns)" and "Occasional (every 8-10 turns)". This should be a binary question with "Memory" as the header.
6. Communication style: "Terse or detailed responses?"
7. Any specific rules: "Rules I should always follow?"

Why these matter:

• Identity lets you correlate git history to the user (their commits, PRs, coding style)
• Related repos provide crucial context (many projects span multiple repos)
• Historical sessions from Claude Code/Codex can reveal preferences, communication style, and project knowledge — but processing them is expensive (parallel subagents, multiple LLM calls), so always ask first
• Workflow/communication style should be stored in system/human/prefs/
• Rules go in system/persona/

What NOT to ask

• Things you can find by reading files ("What's your test framework?")
• "What kind of work do you do? Reviewing PRs vs writing code?" - obvious from git log, most devs do everything
• Permission for obvious actions - just do them
• Questions one at a time - bundle them (but don't exhaust the user with too many questions at once)

During execution, be autonomous. Make reasonable choices and proceed.

Memory File Strategy

Hierarchical Organization (MANDATORY with Memory Filesystem)

With memory filesystem enabled, you MUST organize memory as a deeply nested file hierarchy using bash commands:

NEVER create flat files like project-overview.md, project-commands.md. Instead, create deeply nested structures with / naming:

# Create the hierarchy
mkdir -p ~/.letta/agents/<agent-id>/memory/system/project/tooling
mkdir -p ~/.letta/agents/<agent-id>/memory/system/human/prefs

# Files will be:
# system/project/overview.md
# system/project/commands.md
# system/project/tooling/testing.md
# system/human/identity.md
# system/human/prefs/communication.md

Naming convention (MANDATORY):

• Every new file MUST use / naming - no flat files allowed
• Use / for hierarchy: project/tooling/testing (not project-tooling-testing)
• File path determines the memory label: system/project/overview.md → label project/overview
• Keep files small and focused (~40 lines max)
• Use descriptive frontmatter — the description field helps your future self understand each file's purpose

Checkpoint before proceeding: Count your proposed files. If you have fewer than 15 files, go back and split more aggressively.

Benefits:

• More scannable and maintainable
• Easier to share specific subtrees with other agents
• Natural progressive disclosure (load parent, then drill into children)
• Works like a file system you're familiar with

Split Aggressively - Target 15-25 Files

Don't create monolithic files. Your goal is 15-25 total files. Be aggressive about splitting:

Split when:

• A file has 40+ lines (lower threshold than typical)
• A file has 2+ distinct concepts (not 3+, be aggressive)
• A section could stand alone as its own file
• You can name the extracted content with a clear / path

If a file is getting long (>40 lines), split it:

Without memory filesystem (flat naming - acceptable but not ideal):

• project-overview: High-level description, tech stack, repo links
• project-commands: Build, test, lint, dev commands
• project-conventions: Commit style, PR process, code style
• project-architecture: Directory structure, key modules
• project-gotchas: Footguns, things to watch out for

With memory filesystem (MANDATORY hierarchical naming with /):

• project/overview: High-level description, tech stack, repo links
• project/commands: Build, test, lint, dev commands
• project/conventions: Commit style, PR process, code style
• project/architecture: Directory structure, key modules
• project/gotchas: Footguns, things to watch out for
• Must further nest: project/tooling/testing, project/tooling/linting, project/tooling/bun
• Target 15-25 files total - if commands is long, split into project/commands/dev, project/commands/build, etc.

This makes memory more scannable and easier to update and share with other agents.

Update Memory Incrementally

For deep research: Update memory as you go, not all at once at the end.

Why this matters:

• Deep research can take many turns and millions of tokens
• Context windows overflow and trigger rolling summaries
• If you wait until the end to write memory, you may lose important details
• Write findings to memory files as you discover them

Good pattern:

1. Create file structure early (even with placeholder content)
2. Update files after each research phase
3. Refine and consolidate at the end

There's no reason to wait until you "know everything" to write memory. Treat your memory files as a living scratchpad.

Initialize ALL Relevant Blocks

Don't just update a single memory file. Based on your upfront questions, also update:

• human: Store the user's identity, workflow preferences, communication style
• persona: Store rules the user wants you to follow, behavioral adaptations
• project/*: Split project info across multiple focused files

And add memory files that you think make sense to add (e.g., project/architecture, project/conventions, project/gotchas, or splitting human/ into more focused files, or separate files for multiple users).

Your Task

1. Check memory filesystem status: Look for the memory_filesystem section in your system prompt to confirm the filesystem is enabled.

2. Check for historical session data: Run ls ~/.claude/history.jsonl ~/.codex/history.jsonl 2>/dev/null to see if Claude Code or Codex history exists. You need this result BEFORE asking upfront questions so you know whether to include the history question.

3. Ask upfront questions: Use AskUserQuestion with the recommended questions above (bundled together). This is critical - don't skip it. If history data exists (from step 2), you MUST include the historical sessions question.

4. Inspect existing memory:

   • If memfs enabled: Use ls -la ~/.letta/agents/<agent-id>/memory/system/ to see the file structure
   • Otherwise: Use memory tools to inspect existing files
   • Analyze what exists and what needs improvement

5. Identify the user: From git logs and their answer, figure out who they are and store in system/human/. If relevant, ask questions to gather information about their preferences that will help you be a useful assistant to them.

6. Update human/persona early: Based on answers, update your memory files eagerly before diving into project research. You can always change them as you go, you're not locked into any memory configuration.

7. Research the project: Explore based on chosen depth. Use your TODO or plan tool to create a systematic research plan.

8. Historical session analysis (if approved): If the user approved Claude Code / Codex history analysis in step 3, follow the Historical Session Analysis section below. This launches parallel subagents to process history data and synthesize findings into memory. Skip this step if the user chose "Skip".

9. Create/update memory structure (can happen incrementally alongside steps 7-8):

   • With memfs enabled: Create a deeply hierarchical file structure using bash commands
     • Use mkdir -p to create subdirectories (2-3 levels deep)
     • Create .md files for memory files using / naming
     • Target 15-25 total files - be aggressive about splitting
     • Use nested paths like project/tooling/testing.md (never flat like project-testing.md)
     • Every new file MUST be nested under a parent using /
     • Every new file MUST be nested under a parent using /
   • Without memfs: Use memory tools to create/update files with hierarchical naming
   • Don't wait until the end - write findings as you go

   Checkpoint verification:

   • After creating files, count them: ls ~/.letta/agents/<agent-id>/memory/system/ | wc -l
   • If count < 15, you haven't split enough - go back and split more
   • Check maximum depth: find ~/.letta/agents/<agent-id>/memory/system/ -type f | awk -F/ '{print NF}' | sort -n | tail -1
   • Should be 2-3 levels deep minimum

10. Organize incrementally:

• Start with a basic structure
• Add detail as you research
• Refine organization as patterns emerge
• Split large files into smaller, focused ones

11. Reflect and review: See "Reflection Phase" below - this is critical for deep research.

12. Ask user if done: Check if they're satisfied or want you to continue refining.

13. Push memory: Once the user is satisfied, commit and push your memory repo so changes are synced to the server.

Reflection Phase (Critical for Deep Research)

Before finishing, you MUST do a reflection step. Your memory files are visible to you in your system prompt right now. Look at them carefully and ask yourself:

1. File count check:

   • Count your memory files: ls ~/.letta/agents/<agent-id>/memory/system/ | wc -l
   • Do you have 15-25 files? If not, you haven't split enough
   • Too few files means they're too large - split more aggressively

2. Hierarchy check:

   • Are ALL new files using / naming? (e.g., project/tooling/bun.md)
   • Do you have 2-3 levels of nesting minimum?
   • Are there any flat files like project-commands.md? These should be nested

3. Redundancy check: Are there files with overlapping content? Either literally overlapping (due to errors while editing), or semantically/conceptually overlapping?

4. Completeness check: Did you actually update ALL relevant files? For example:

   • Did you update human with the user's identity and preferences?
   • Did you update persona with behavioral rules they expressed?
   • Or did you only update project files and forget the rest?

5. Quality check: Are there typos, formatting issues, or unclear frontmatter descriptions?

6. Structure check: Would this make sense to your future self? Is anything missing? Is anything redundant?

After reflection, fix any issues you found. Then ask the user:

"I've completed the initialization. Here's a brief summary of what I set up: [summary]. Should I continue refining, or is this good to proceed?"

This gives the user a chance to provide feedback or ask for adjustments before you finish.

Working with Memory Filesystem (Practical Guide)

Here's how to work with the memory filesystem during initialization:

Inspecting Current Structure

# See what memory files currently exist
ls -la ~/.letta/agents/<agent-id>/memory/system/

# Check the tree structure
tree ~/.letta/agents/<agent-id>/memory/system/

# Read existing memory files
cat ~/.letta/agents/<agent-id>/memory/system/persona.md

Creating Hierarchical Structure (MANDATORY)

Good examples (nested with /): ✅ project/overview.md ✅ project/tooling/bun.md
✅ project/tooling/testing.md ✅ human/prefs/communication.md ✅ persona/behavior/tone.md

Bad examples (flat naming - NEVER do this): ❌ project-overview.md (flat, not nested) ❌ bun.md (orphan file, no parent) ❌ project_testing.md (underscore instead of /)

# Create deeply nested directory structure (2-3 levels)
mkdir -p ~/.letta/agents/<agent-id>/memory/system/project/{tooling,architecture,conventions}
mkdir -p ~/.letta/agents/<agent-id>/memory/system/human/prefs
mkdir -p ~/.letta/agents/<agent-id>/memory/system/persona/behavior

# Create memory files using Write tool - ALL files must be nested
Write({
  file_path: "~/.letta/agents/<agent-id>/memory/system/project/overview.md",
  content: "## Project Overview\n\n..."
})

Write({
  file_path: "~/.letta/agents/<agent-id>/memory/system/project/tooling/testing.md",
  content: "## Testing Setup\n\n..."
})

Write({
  file_path: "~/.letta/agents/<agent-id>/memory/system/project/tooling/bun.md",
  content: "## Bun Configuration\n\n..."
})

Organizing Existing Files

# If you have flat files that should be hierarchical
mv ~/.letta/agents/<agent-id>/memory/system/project-tooling.md \
   ~/.letta/agents/<agent-id>/memory/system/project/tooling.md

# Create subdirectories as needed
mkdir -p ~/.letta/agents/<agent-id>/memory/system/project/tooling
mv ~/.letta/agents/<agent-id>/memory/system/project/tooling.md \
   ~/.letta/agents/<agent-id>/memory/system/project/tooling/overview.md

Final Checklist (Verify Before Submitting)

Before you tell the user you're done, confirm:

• File count is 15-25 — Count your files with ls ~/.letta/agents/<agent-id>/memory/system/ | wc -l. If < 15, split more.
• All new files use / naming — No flat files like my_notes.md or project-commands.md
• Hierarchy is 2-3 levels deep — e.g., project/tooling/bun.md, not just project.md
• No file exceeds ~40 lines — Split larger files
• Each file has one concept — If 2+ topics, split into 2+ files
• Every file has real content — No empty or pointer-only files
• Verify sync: After creating files, check they appear in your memory files

If you have fewer than 15 files, you haven't split enough. Go back and split more.

Best Practices

1. Check memfs status first: Look for memory_filesystem section in your system prompt
2. Start with directories: Create the directory structure before populating files
3. Use short paths: Aim for 2-3 levels (e.g., project/tooling/testing, not project/dev/tools/testing/setup)
4. Keep files focused: Each file should cover one concept (~40 lines max)
5. Every file should have real content — no empty or pointer-only files
6. Be aggressive about splitting: If in doubt, split. Too many small files is better than too few large ones.

Remember: Good memory management is an investment. The effort you put into organizing your memory now will pay dividends as you work with this user over time.

Historical Session Analysis (Optional)

This section runs only if the user approved during upfront questions. It uses parallel history-analyzer subagents to process Claude Code and/or Codex history into memory. The subagents automatically have the migrating-from-codex-and-claude-code skill loaded for data access.

Architecture: Parallel worker subagents each process a slice of the history data (on their own git branch in the memory repo), then a synthesis agent merges all branches and updates memory. The workers serve the same goals as the rest of this initialization skill — understanding the user, their preferences, communication style, project context, and anything that makes the agent more effective. Split data however makes sense — by date range, by source (Claude vs Codex), or both.

Prerequisites:

• letta.js must be built (bun run build) — subagents spawn via this binary
• Use subagent_type: "history-analyzer" — cheaper model (sonnet), has bypassPermissions, creates its own worktree

Step 1: Detect Data, Plan Splits, and Pre-split Files

ls ~/.claude/history.jsonl ~/.codex/history.jsonl 2>/dev/null
wc -l ~/.claude/history.jsonl ~/.codex/history.jsonl 2>/dev/null

Split the data across multiple workers for parallel processing — the more workers, the faster it completes. Use 2-4+ workers depending on data volume.

Pre-split the JSONL files by line count so each worker reads only its chunk. This is simpler than date-based splitting and guarantees evenly-sized chunks:

SPLIT_DIR=/tmp/history-splits
mkdir -p "$SPLIT_DIR"
NUM_WORKERS=3  # adjust based on data volume

# Split Claude history into even chunks
LINES=$(wc -l < ~/.claude/history.jsonl)
CHUNK_SIZE=$(( LINES / NUM_WORKERS + 1 ))
split -l $CHUNK_SIZE ~/.claude/history.jsonl "$SPLIT_DIR/claude-"

# Split Codex history (if it exists and is large enough to warrant splitting)
if [ -f ~/.codex/history.jsonl ]; then
  LINES=$(wc -l < ~/.codex/history.jsonl)
  if [ "$LINES" -gt 100 ]; then
    CHUNK_SIZE=$(( LINES / NUM_WORKERS + 1 ))
    split -l $CHUNK_SIZE ~/.codex/history.jsonl "$SPLIT_DIR/codex-"
  else
    cp ~/.codex/history.jsonl "$SPLIT_DIR/codex-aa"
  fi
fi

# Rename to .jsonl for clarity
for f in "$SPLIT_DIR"/*; do mv "$f" "$f.jsonl" 2>/dev/null; done

# Verify even splits
wc -l "$SPLIT_DIR"/*.jsonl

This is critical for performance — workers read a small pre-filtered file instead of scanning the full history on every query.

Step 2: Launch Workers in Parallel

Send all Task calls in a single message. Each worker creates its own worktree, reads its pre-split chunk, directly updates memory files, and commits. Workers do NOT merge.

Task({
  subagent_type: "history-analyzer",
  description: "Process chunk [N] of [SOURCE] history",
  prompt: `## Assignment
- **Memory dir**: [MEMORY_DIR]
- **History chunk**: /tmp/history-splits/[claude-aa.jsonl | codex-aa.jsonl]
- **Source format**: [Claude (.timestamp ms, .display) | Codex (.ts seconds, .text)]
- **Session files**: [~/.claude/projects/ | ~/.codex/sessions/]
`
})

Step 3: Merge Worker Branches and Curate Memory (you do this yourself)

After all workers complete, you (the main agent) merge their branches back into main and then review, curate, and reorganize the resulting memory. This is critical — workers produce raw output that needs editorial judgment.

3a. Merge branches:

cd [MEMORY_DIR]
for branch in $(git branch | grep migration-); do
  git merge $branch --no-edit -m "merge: $branch"
done

If there are merge conflicts, read both versions and keep the most complete content. Resolve them yourself — it's just text.

3b. Review and curate merged memory:

After merging, read every file in system/ and apply editorial judgment:

• Only high-signal, actionable information belongs in system/ — this is rendered in-context every turn and directly affects token cost and response quality
• Move supplementary/reference content to reference/ — detailed history, evidence, examples, verbose context that's useful but not needed every turn
• Deduplicate across workers — multiple workers may have written overlapping or redundant content to the same files. Consolidate into clean, non-repetitive content
• Reformat for scannability — bullet points, short lines, no walls of text. Your future self needs to parse this instantly
• Delete low-value content — if something isn't clearly useful for day-to-day work, remove it. Less is more in system/

3c. Reorganize file structure if needed:

Workers may have created files that don't fit the ideal hierarchy, or put too much into system/. Fix this:

• Split oversized files (>40 lines) into focused sub-files
• Move reference-quality content (detailed history, background context, evidence trails) to reference/
• Ensure system/ contains only what you genuinely need in-context: identity, active preferences, current project context, behavioral rules, gotchas
• Merge near-duplicate files that cover the same topic

Rule of thumb: If removing a file from system/ wouldn't hurt your next 10 responses, it belongs in reference/.

3d. Clean up worktrees and branches:

for w in $(dirname [MEMORY_DIR])/memory-worktrees/migration-*; do
  git worktree remove "$w" 2>/dev/null
done
git branch -d $(git branch | grep migration-)
git push

Troubleshooting

Problem	Cause	Fix
Subagent exits with code null, 0 tool uses	letta.js not built	Run bun run build
Subagent hangs on "Tool requires approval"	Wrong subagent type	Use subagent_type: "history-analyzer" (workers) or "memory" (synthesis)
Merge conflict during synthesis	Workers touched overlapping files	Resolve by checking git log for context
Auth fails on push ("repository not found")	Credential helper broken	Use http.extraHeader (see syncing-memory-filesystem skill)