MnemosyneEcto (mnemosyne_ecto v0.2.0)

Database-agnostic Ecto backend for the Mnemosyne agentic memory library.

Provides MnemosyneEcto.Backend, a Mnemosyne.GraphBackend implementation for PostgreSQL with pgvector or SQLite with sqlite-vec. The configured Ecto adapter selects the engine; MnemosyneEcto.Adapter contains engine-specific vector, migration, and extension behavior.

Graph nodes, node metadata, and permanent ingestion records use three dynamically prefixed tables. Ingestion records are scoped by tenant, repository, and source ID. They preserve a payload digest, fingerprint version, ordered node IDs, and microsecond stored timestamp without a graph foreign key, so receipts survive graph-node deletion.

Trajectory ingestion

Configure MnemosyneEcto.Backend normally, then submit caller-owned complete trajectories through Mnemosyne.ingest/3. It blocks until graph changes and the durable receipt are stored, or returns an error. The returned node IDs are immediately queryable. Equal retries return the exact original receipt; conflicting source reuse returns an ingestion error. Database uniqueness selects the concurrent winner, making receipts durable across restarts and independent across tenants and repositories.

PostgreSQL and SQLite have equivalent observable behavior. SQLite retries one complete rolled-back transaction only for Exqlite's locked "Database busy" error; a second or nonmatching error is a StorageError.

MnemosyneEcto.Telemetry documents the get_ingestion and commit_ingestion spans, including correlation, status, and numeric count measurements.

Choosing a database

Configure your own Ecto.Repo with the desired adapter:

# PostgreSQL
defmodule MyApp.Repo do
  use Ecto.Repo, otp_app: :my_app, adapter: Ecto.Adapters.Postgres
end

# SQLite (load the sqlite-vec extension on every connection)
defmodule MyApp.Repo do
  use Ecto.Repo, otp_app: :my_app, adapter: Ecto.Adapters.SQLite3
end

config :my_app, MyApp.Repo, load_extensions: [SqliteVec.path()]

Setup

{Mnemosyne.Supervisor,
  config: config,
  llm: MyApp.LLM,
  embedding: MyApp.Embedding,
  backend: {MnemosyneEcto.Backend, repo: MyApp.Repo}}

# The configured backend defaults tenant_id to "default".
Mnemosyne.open_repo("my-project")

Mnemosyne.open_repo("my-project",
  backend: {MnemosyneEcto.Backend, repo: MyApp.Repo, tenant_id: "org-123"}
)

Migrations

A V1 migration creates all three tables for either adapter:

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

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

This is a clean break. Drop and recreate MnemosyneEcto tables or databases from pre-ingestion versions. V1 was rewritten in place: current_version/0 remains 1, with no V2 upgrade, data conversion, session compatibility, or dual read/write path.