Quickstart

Install

pip install continuum-ai

The wheel bundles the compiled engine. The import path is continuum; the runtime classes live in continuum._native.

From a checkout, with uv:

git clone https://github.com/rithulkamesh/continuum
cd continuum
CMAKE_BUILD_PARALLEL_LEVEL=2 uv sync --all-extras

The parallelism cap matters: the extension links libtorch with LTO, and an unbounded build can exhaust memory on a 16 GB machine.

Run the examples

Three scripts under examples/ each demonstrate one capability. They use the FakeLLM backend, so output is deterministic and safe to run in CI.

PYTHONPATH=python python examples/01_reuse_stack.py       # every reuse tier, one run
PYTHONPATH=python python examples/02_durable_agent.py     # checkpoint, crash, resume
PYTHONPATH=python python examples/03_time_travel_fork.py  # rewind, edit, replay

Add --trace to 01_reuse_stack.py to see which tier answered each call.

A minimal program

The frontend traces a decorated Python function to the canonical IR on its first call, then runs it through the engine.

from continuum import program

@program
def pipeline(question: str):
    # token and tensor steps recorded here become IR nodes
    ...

pipeline("what changed in the last release?")

Tunable values are declared with Param and searched by Optimizer; parameter discovery uses the Module base class.

Next