unCAGd Skill#

Use this skill when the user is working on a repo, software project, agent workflow, plugin, MCP server, or multi-step implementation where decisions may matter later.

When to retrieve memory#

Before planning or editing code, call cag.retrieve when:

• The user references prior decisions.
• The task is part of an ongoing project.
• The user asks to continue, refine, implement, or fix something.
• The task involves architecture, naming, APIs, tests, security, or deployment.

Use k=5 unless the user explicitly asks for a broader review. Set useEmbeddings=true only as a secondary ranker when lexical retrieval is weak.

When to capture a candidate#

Call cag.capture_candidate when the conversation produces a durable decision such as:

• Architecture choice.
• Naming convention.
• API contract.
• Security constraint.
• Testing strategy.
• Deployment assumption.
• Project-specific behavior.

Captured candidates are not durable until validated.

When to validate#

Call cag.validate_memory only when:

• The user explicitly approves the decision.
• A test passes and the decision is tied to that test.
• A PR/commit/doc source is provided.
• The user says to remember the decision for the project.

Important#

Do not store raw assistant guesses as validated memory. Do not use CAG memory as a replacement for reading the codebase. Use retrieved decisions as constraints, not as proof of current code state.

Contradiction workflow#

When contradictions are detected:

• Use cag.resolve_contradiction with keep_existing if the prior decision stands.
• Use supersede_existing if policy changed and old validated memories should be deprecated.
• Use create_candidate if the new decision needs review first.

Compression and portability#

• Use cag.compress_memory to consolidate old validated decisions into one compact, validated summary.
• Use cag.export_memory with format=memory_md for portable handoff docs.
• Use cag.import_memory to rehydrate memory in another workspace.