Database-agnostic migrations for MnemosyneEcto tables and indexes.

Write one migration in your application. MnemosyneEcto.Adapter emits the correct DDL for PostgreSQL with pgvector or SQLite with sqlite-vec. V1 creates all three dynamically prefixed tables: graph nodes, node metadata, and permanent ingestion records.

Usage

defmodule MyApp.Repo.Migrations.AddMnemosyne do
  use Ecto.Migration

  def up, do: MnemosyneEcto.Migrations.up(version: 1, embedding_dimensions: 1536)
  def down, do: MnemosyneEcto.Migrations.down(version: 1)
end

Clean-break reset

V1 was rewritten for durable ingestion records. Existing MnemosyneEcto tables or databases from pre-ingestion versions must be dropped and recreated. The current version remains 1; there is no V2 upgrade, data conversion, session compatibility, or dual read/write path.

Options

• :version - the target migration version (defaults to 1)
• :embedding_dimensions - required for V1; dimensionality of embedding vectors
• :index_type - PostgreSQL only: :hnsw (default) or :ivfflat
• :hnsw_m - PostgreSQL only: max number of connections per HNSW layer
• :hnsw_ef_construction - PostgreSQL only: dynamic candidate list size for HNSW
• :ivfflat_lists - PostgreSQL only: number of inverted lists for IVFFlat
• :prefix - table name prefix, defaults to "mnemosyne_"

SQLite uses a brute-force vec_distance_cosine scan, so PostgreSQL-only index options are ignored for SQLite repos.