[Krokz]

Summary

• define_schema with only preset="memory" (MCP, REST, service) adopts v1 of an optional memory preset: Decision, Insight, Task, Question, each with STRING id key and title, body, status, scope.
• Adoption is additive and repeatable. It creates missing tables and adds missing properties with typed defaults. It never drops, renames or retypes. Key/type mismatches, relationship-name clashes and near-duplicate tables are reported and left unchanged; allow_similar creates beside a near-duplicate.
• The adopted version is recorded in _grag_meta (preset.memory.version). Adoption refuses a newer recorded version. Non-additive migration steps run once when crossing their version and must be idempotent, because export does not carry _grag_meta.
• An existing table's FTS index is rebuilt after commit when adoption adds text columns. DROP_FTS_INDEX only runs in auto-transaction mode.
• CLI: grag memory adopt [--allow-similar] [--json] and grag init --memory-preset. The client refuses owners without the new memory_preset health capability before sending anything.
• The schema revision and describe_schema output are unchanged. The preset report appears only in the define_schema footer.
• Docs cover the architecture principle, the schema reference section, CLI/MCP references, the memory guide and all four skill-reference copies.

Surface cost

define_schema advertised schema: 1,490 → 1,658 B (+168 B), from the preset param, optional table lists and one docstring sentence. Instructions are unchanged. The discovery budget test still passes.

Test plan

• tests/test_presets.py: fresh adoption, rerun, FTS coverage across reopen (mutation-checked), conflicts re-reported, downgrade refusal, migration crossing, validation, MCP, CLI, init, old-owner refusal
• Full suite, ruff, mypy locally

🤖 Generated with Claude Code