Memory Types

Memory files can declare a type: field in their frontmatter to signal their purpose and content category. This helps the selection algorithms prioritize and filter memories during recall.

Standard types

Claude Code recognizes four standard memory types:

type: user

Personal insights, conventions, and learned practices specific to the user’s workflow.

Examples:

• “I prefer using async/await over promises”
• “Our team uses Conventional Commits for all PRs”
• “I always check X before deploying to production”

Scope: User-global; applicable across all projects.

type: feedback

Captured user feedback, feature requests, and bug reports from stakeholders or automated systems.

Examples:

• “User reported lag in search feature on slow networks”
• “Analytics show 20% drop in onboarding completion after v2 release”
• “Feature request: Dark mode support in the dashboard”

Scope: Can be user-global or project-specific depending on source.

type: project

Project-specific findings, architecture decisions, code patterns, and domain knowledge.

Examples:

• “The auth service uses JWT with RS256 signature”
• “Database migrations must support zero-downtime deployments”
• “All API endpoints return a meta object with request ID and timestamp”

Scope: Project-local; most relevant when in that project context.

type: reference

External links, documentation excerpts, API specifications, and reference materials.

Examples:

• Links to relevant RFCs, standards, or third-party API docs
• Excerpts from official documentation
• Saved runbooks or troubleshooting guides

Scope: Can be user-global for widely-used references or project-local for project-specific specs.

Untyped files

Files without a type: field are still scanned and selected, but may be weighted lower by the Sonnet selector during recall. It is recommended to add a type if the file will be in a memory directory.

Frontmatter example

---
type: project
description: Architecture decisions for the API layer
---

# API Architecture

## Request/response format
[...]

Type-based narrowing

The selection algorithm may narrow by type when the current context suggests a particular focus:

• Active project context → prefer project type
• Feature discussion → prefer feedback type
• Personal workflows → prefer user type
• Reference queries → prefer reference type

However, type is a hint, not a hard filter. All files are still eligible for selection if the Sonnet selector deems them relevant.

---

← Back to Memory/README.md