SQL Memory for agents. The third facet of the Memory primitive (v1.2.0)

Loomcycle v1.2.0 introduces SQL Memory, a third facet of the Memory primitive, enabling authorized agents to run arbitrary SQL against per-scope databases. The update includes two new ops (sql_exec, sql_query) and three transaction ops (sql_begin, sql_commit, sql_rollback), supporting SQLite and PostgreSQL with scoped access, audit, and vector column integration. This release is additive and off-by-default, with no breaking changes.

loomcycle v1.2.0 ships RFC AA Phases 1 through 3g: SQL Memory. A third facet of the Memory primitive. Authorized agents run arbitrary SQL against a per-scope database the runtime hosts, isolated from the main loomcycle store. Two new ops sql exec, sql query plus three for transactions sql begin, sql commit, sql rollback on the existing Memory tool. Two tiers: sqlite file-per-scope, statement-allowlist hardened — the default modernc.org/sqlite driver has no authorizer, so the primary defense is a Go-layer parsed-statement validator and postgres schema-per-scope, per-scope least-privilege LOGIN role with search path pinned . Three scopes: durable agent/user tenant-keyed, persist across runs plus ephemeral run one DB per spawn tree, dropped at completion . Default-deny sql scopes ACL per agent. Per-statement timeout, per-scope byte quota, row cap, full audit redacted via RFC Z . Phase 3a explicit transactions with runtime-managed cleanup TTL reaper for abandoned transactions . Phase 3b nested via SAVEPOINT LIFO, depth-16 cap . Phase 3c vector columns inside agents' own tables on postgres — semantic KNN AND structured filters in one query, with server-side $embed substitution so multi-KB vectors never round-trip through the LLM. Phase 3d + 3f.3 durable-scope GC TTL + size-budget variants, off by default + lossy by contract . Phase 3e + 3f.2 snapshot integration with per-scope cap. Phase 3g read-only shared schemas — operator-blessed reference data on every scope role's search path. exp9 demonstrates the shape: prime-stream → SQL table in user scope → channel ping → validator agent on a different run reads the same table, validates each prime, writes verdicts to a second table. Additive + off-by-default — no breaking changes, no new wire RPCs; TS + Python adapters unchanged at 1.1.1.