Install

Benchmark first

Start with one slice each from BEIR, FEVER, and LongMemEval. They expose retrieval, evidence, updates, temporal reasoning, and abstention. Record quality, latency, tokens, storage, and corpus revision from the first prototype; the same cases become regression gates as components change.

Python 3.11–3.13 is supported. NumPy is the only runtime dependency.

terminal

pip install mari-components

# Optional runtime adapters
pip install 'mari-components[openai-agents]'
pip install 'mari-components[langchain]'

How it works

The base wheel installs the domain, connector, synchronization, retrieval, knowledge, trajectory, and verification modules without a model SDK or database client. Extras add adapter imports only; applications still inject model calls, HTTP transport, persistence, clocks, and authorization decisions. This is package behavior, not a research-derived algorithm.