Create Memory

Generated Claude Code skill — do not edit manually

This complete skill and its bundled resources are generated from agent-plugins/skills/create-memory/SKILL.md. Manual changes in this directory will be overwritten the next time the generator runs. Edit the canonical skill instead.

Create Memory

Create the smallest useful project-memory baseline without crossing authorization boundaries.

Load the Contract and Resolve Roots

1. Read references/memory-contract.md completely before proposing files.
2. Resolve the project root and memory root from the user's arguments. Otherwise, consider only plausible roots inside the current project.
3. Ask for explicit consent before reading or writing a global store, sibling project, unrelated repository, or any root outside the selected project. Relevance alone is not authorization.
4. If several roots are plausible, show the candidates and ask the user to choose. Never merge stores silently.
5. Read any existing MEMORY.md and local memory files. If a useful baseline already exists, route durable session updates to update-memory and source imports to ingest-memory.

Build a Durable Candidate Set

Inspect only relevant repository instructions, README and architecture material, manifests, important entry points, schemas, integration boundaries, and available read-only branch or history evidence. Exclude dependencies, generated output, large data, credentials, secrets, private personal data, and unrelated memories.

Capture only stable, non-obvious knowledge that will change future work:

• project: architecture, workflows, boundaries, conventions, integrations, and adopted decisions
• feedback: recurring user preferences, corrections, and explicit working agreements
• reference: stable technical facts future work needs

Skip temporary status, unfinished experiments, raw logs or conversations, speculative ideas, facts already enforced by code or instructions, and details that are cheap to rediscover.

Preview and Write

1. Present a compact plan with the selected roots, proposed files, types, evidence, relationships, and skipped candidates.
2. Ask for confirmation when the root was inferred, any cross-root access is involved, more than three files would be created, an index would be replaced, or existing content would change non-additively.
3. Re-read the target files immediately before editing.
4. Follow the contract exactly. Use its versioned frontmatter, structured provenance, local symmetric related filenames, ISO dates, and exact Markdown index-entry grammar.
5. Create focused files with one durable topic each. Preserve manual index entries and unrelated content. Do not delete stale material without an explicit cleanup request.
6. Compare every proposed filename, name, description, source identity, and distinctive claim with existing memory. Merge or report duplicates and preserve unresolved conflicts.

Validate and Report

Run the bundled read-only validator after writing:

python scripts/validate_memory.py <memory-root>

Fix only errors introduced within the approved plan, then rerun validation. Report selected roots, files created or changed, evidence used, skipped candidates, duplicate or conflict decisions, validation results, and unresolved limitations. Recommend audit-memory after multi-file creation.