z3rno-core 0.21.3

Core memory engine for Z3rno: PostgreSQL schema, SQLAlchemy models, Alembic migrations, and the store/recall/forget/audit library functions.

z3rno-core

Core memory engine for Z3rno -- PostgreSQL schema, SQLAlchemy models, Alembic migrations, and the store/recall/forget/audit library.

Features

• 7 relational tables -- tenants, API keys, agents, memories, memory relationships, lifecycle policies, audit log
• 15 Alembic migrations -- fully versioned schema evolution
• Row-Level Security (RLS) -- multi-tenant isolation at the database level
• SCD Type 2 versioning -- temporal history for every memory mutation
• pgvector HNSW indexing -- fast approximate nearest-neighbor vector search
• Apache AGE graph layer -- entity and relationship traversal via Cypher queries
• Hash-chained audit log -- tamper-evident record of every store, recall, forget, and update

Quickstart

Run migrations

# Set your database URL
export DATABASE_URL="postgresql://z3rno:password@localhost:5432/z3rno"

# Install
pip install -e .

# Apply all migrations
alembic upgrade head

Use the engine

from sqlalchemy.ext.asyncio import create_async_engine, async_sessionmaker
from z3rno_core.engine import store, recall, forget, audit

engine = create_async_engine("postgresql+asyncpg://z3rno:password@localhost/z3rno")
Session = async_sessionmaker(engine)

async with Session() as db:
    # Store a memory
    memory = await store(db, org_id=org_id, agent_id="agent-1", content="User prefers dark mode")

    # Recall by semantic similarity
    results = await recall(db, org_id=org_id, agent_id="agent-1", query="user preferences", top_k=5)

    # Soft-delete a memory
    await forget(db, org_id=org_id, memory_id=memory.id)

    # Query the audit trail
    entries = await audit(db, org_id=org_id, agent_id="agent-1")

For a detailed step-by-step setup, see QUICKSTART.md.

Full documentation: astron-bb4261fd.mintlify.app

Architecture

The schema follows a strict dependency order:

tenants -> api_keys -> agents -> memories -> memory_relationships
                                    |
                                    +-> lifecycle_policies
                                    +-> audit_log

All engine functions operate within a single async SQLAlchemy session and respect RLS via SET LOCAL on the org context.

Documentation

• SCHEMA.md -- full table and column reference
• MULTI_TENANCY.md -- RLS design and tenant isolation
• Architecture Decision Records -- design rationale

Development

uv sync --dev
uv run ruff check .
uv run mypy .
uv run pytest

See CONTRIBUTING.md for the full workflow.

License

Apache 2.0 -- see LICENSE.