Memory Init

Memory Init

Scaffold lightweight, project-local memory files that future agents can read and maintain.

Use this skill to initialize durable project knowledge, not to store one-off task notes. The default target is .agents/memories/, with AGENTS.md guidance for Codex-oriented future use.

Workflow

1. Confirm the target project root.
2. Inspect whether the target memory directory and agent instruction file already exist.
3. Use .agents/memories/ unless the user specifies another memory directory.
4. Run scripts/init_project_memory.py.
5. Review the generated or changed files.
6. Report the memory directory and whether the agent instruction file was created or updated.

Command

From any location:

python3 <skill-dir>/scripts/init_project_memory.py /path/to/project

With a custom memory directory:

python3 <skill-dir>/scripts/init_project_memory.py /path/to/project --memory-dir docs/agent-memory

To create memory files without touching the agent instruction file:

python3 <skill-dir>/scripts/init_project_memory.py /path/to/project --skip-agent

With an alternate agent instruction file:

python3 <skill-dir>/scripts/init_project_memory.py /path/to/project --agent-file CLAUDE.md

Generated Files

The scaffold creates missing files only. Existing memory files are preserved.

• README.md: explains the memory system, what belongs there, and what must not be stored.
• index.md: routes future agents to the relevant memory files.
• project.md: stores stable project conventions and durable context.
• workflows.md: stores repeatable commands, release steps, CI workflows, and maintenance procedures.
• lessons.md: stores reusable lessons, pitfalls, and verified fixes.

AGENTS.md Guidance

By default, the script creates or updates AGENTS.md in the project root. It inserts a concise ## Project Memory section that tells future agents to:

• read index.md before non-trivial work;
• open only task-relevant memory files;
• update memories when the user requests memory maintenance or when the current task explicitly includes maintaining reusable project knowledge;
• recommend a memory update instead of editing memory files when memory maintenance is outside the current task scope;
• avoid one-off task notes, temporary debugging details, secrets, credentials, and conversation-specific context;
• keep index.md aligned when memory files are added, removed, renamed, or materially changed.

If an existing managed block is present, update it instead of appending a duplicate. If only an unmarked ## Project Memory / ## Project Memories section exists, preserve the user-authored section and append a managed block.

Memory Standard

Treat memory as durable project knowledge. Add or update entries when the user asks for memory maintenance or the task explicitly includes maintaining reusable project context.

Good memory candidates:

• stable repository conventions;
• commands or workflows that were verified locally;
• recurring environment constraints;
• cross-task architectural or maintenance decisions;
• reusable debugging lessons.

Do not store:

• secrets, credentials, tokens, or private keys;
• raw conversation logs;
• single-task progress notes;
• speculative conclusions that have not been verified;
• noisy command output that is not reusable.