ADR 0001: Start with a local Python and SQLite kernel

• Status: accepted for Prototype A
• Date: 2026-07-21

Context#

The repository is greenfield. The first risk is architectural validity, not distributed scale. Requiring a vector database, graph database, hosted model, or API key would make the earliest results slower to reproduce and would mix infrastructure effects with the memory design.

Decision#

Build the deterministic kernel in Python 3.12 using the standard library and SQLite FTS5. Keep storage, extraction, embedding, reranking, and token-counting behind narrow interfaces.

The import package is evicortex. The distribution remains evicortex-research until the PyPI namespace and formal trademark checks are completed; no public package name is reserved.

Consequences#

Benefits:

• one-file local setup and reproducible tests;
• ACID transactions, FTS5, indexes, and a mature query planner;
• no network, model, or paid-service dependency for the baseline;
• easy inspection of invariants and benchmark artifacts.

Costs:

• no dense semantic retrieval in Prototype A;
• graph traversal and rank fusion must be implemented in application code;
• write concurrency and very large deployments will require another adapter;
• SQLite benchmark results cannot be presented as distributed production results.

Exit criteria#

Add or replace storage only after profiling identifies a concrete limit. A replacement must pass the same conformance suite for evidence fidelity, temporal claims, isolation, expansion, and deterministic retrieval traces.