| name | memory-architecture-contract |
| description | Guides Gantry memory architecture work around session digest, continuity, compaction, dreaming, retrieval, source-first ingestion, and cleanup. Use when planning or changing memory services, generated memories, memory APIs, MCP tools, settings, or retrieval behavior. |
Memory Architecture Contract
Use this skill for memory planning, memory service fixes, generated memories,
retrieval claims, compaction handling, and source-first knowledge management.
Required Workflow
- Read relevant docs under
docs/architecture/, docs/decisions/, and current memory service code before choosing a design.
- Prefer digest-first continuity for long sessions; do not reduce memory design to small fact extraction when the task is about session continuity or generated memories.
- Treat compaction summaries as high-signal inputs, not raw durable memory.
- Keep dreaming deterministic when used for maintenance: schema-validated operations, idempotency, rollback or replay safety, and auditability.
- Treat blogs, docs, posts, and files as immutable memory sources feeding reviewed memory, not direct arbitrary
memory_save writes.
- Be precise about retrieval. Do not claim semantic/vector retrieval is active unless the current code and tests prove it.
- Include cleanup across CLI,
settings.yaml, control API, MCP tools, parser/renderer, docs, and tests when replacing memory behavior.
Evidence To Provide
- Memory source, continuity, compaction, dreaming, retrieval, and embedding impact.
- API/CLI/MCP/settings surfaces changed or explicitly unchanged.
- Deterministic validation and rollback/audit plan for generated memory operations.
- Tests or code evidence backing any retrieval capability claim.