Generation Targets

The memories generate command creates native configuration files that each AI tool understands natively. It uses a two-step pipeline to ensure every tool gets the richest possible configuration.

Two-Step Pipeline

Memory Store → .agents/ (canonical) → Tool-specific configs

1. Generate .agents/ — Memories are written to a canonical, tool-agnostic directory structure
2. Adapt per tool — The .agents/ content is translated into each target's native format

This architecture means you have a single source of truth (.agents/) that can generate configurations for any number of tools simultaneously.

The .agents/ Directory

The canonical intermediate format. Generated by memories generate agents (and automatically by other targets):

.agents/
├── instructions.md           # Always-on rules, decisions, facts
├── rules/                    # Path-scoped rules with YAML frontmatter
│   ├── api.md
│   └── testing.md
├── skills/                   # Agent Skills standard (SKILL.md)
│   └── deploy/
│       └── SKILL.md
└── settings.json             # Permissions, hooks, env vars

instructions.md

Global context from non-path-scoped memories of type rule, decision, and fact. Equivalent to what goes into CLAUDE.md and other flat instruction outputs.

rules/*.md

Path-scoped rules. Each file has YAML frontmatter with paths: globs:

---
paths:
  - "src/api/**"
  - "lib/api/**"
---

- Use RESTful naming conventions
- Return consistent error response format

skills/**/SKILL.md

Follows the Agent Skills open standard. Each skill lives in its own category directory with optional supporting files.

settings.json

Agnostic permission, hook, and env configuration. Modeled on Claude Code's settings.json structure, which is the superset most other tools can map from.

Supported Targets

.agents/ (Canonical)

• Path: .agents/
• Format: Structured directory with markdown, YAML frontmatter, and JSON
• The source of truth that all other targets adapt from

Cursor

• Path: .cursor/rules/
• Format: MDC (Markdown with YAML frontmatter)
• instructions.md → .cursor/rules/memories.mdc (with alwaysApply: true)
• rules/*.md → .cursor/rules/{name}.mdc (translates paths: → globs:)
• Skills are copied as-is

Claude Code

• Path: CLAUDE.md + .claude/
• Format: Markdown
• instructions.md → CLAUDE.md
• rules/*.md → .claude/rules/*.md (format is native)
• skills/ → .claude/skills/ (format is identical)
• settings.json → .claude/settings.json (permissions merged)

Agents (Canonical)

• Path: .agents/
• Format: Structured directory
• Used as the canonical intermediate for multi-tool adapters

GitHub Copilot

• Path: .github/copilot-instructions.md
• Format: Markdown (flat file with inline path annotations)

Windsurf

• Path: .windsurf/rules/memories.md
• Format: Markdown
• Has a 6000 character limit — content is truncated at line boundaries if exceeded

Cline

• Path: .clinerules/memories.md
• Format: Markdown

Roo

• Path: .roo/rules/memories.md
• Format: Markdown

Gemini

• Path: GEMINI.md
• Format: Markdown

File Marker

All generated files include a marker comment:

<!-- Generated by memories.sh at 2026-01-15T10:30:00Z -->

This marker serves two purposes:

1. Overwrite protection — memories generate won't overwrite files without this marker unless --force is used
2. Ingest detection — memories ingest skips files with this marker to avoid circular imports

Gitignore Recommendations

After generating files, memories.sh may suggest adding tool-specific directories to .gitignore. Whether you commit generated files is up to you:

• Commit them if you want the rules visible to collaborators who don't use memories.sh
• Gitignore them if all team members use memories.sh and generate locally