Continuum

The AI runtime that never computes the same thing twice, and never loses its place.

Continuum is a C++ execution engine that runs LLM calls and tensor ops as operators in one dataflow graph. Work it has already done is reused at the runtime level, and a running job can be checkpointed to bytes, resumed in another process, or forked from any past step.

pip install continuum-ai

The engine is C++. This site documents the Python surface: the frontend for writing programs, and the runtime classes exposed through continuum._native.

What it does

Reuse. Every call checks four caches in order (exact memo, semantic, shared-prefix KV, warm layer KV) and stops at the first hit. Only a cold call reaches the backend. On a mixed 20-step agent run against a live Azure OpenAI backend this cut tokens sent by 92.5 percent. See The reuse stack.

Durable execution. A checkpoint holds the graph, every computed value, and the KV cache. Write it to a file and a different process reads it and carries the run forward, warm, with deterministic replay. See Durable execution.

One graph for tokens and tensors. LLM calls and tensor ops are scheduled by the same cache-aware interpreter. Swap providers without touching the graph. See Backends and the IR.

A first look

from continuum import DurableAgent

agent = DurableAgent()
agent.begin(["pull the ticket", "reproduce the bug", "draft a fix", "open the PR"])

ckpt = agent.run_until_step(1)              # steps 1 and 2 run now
DurableAgent.inspect(ckpt)                  # {'executed_nodes': 2, 'checkpoint_bytes': ...}

outputs = DurableAgent().resume_from(ckpt)  # a fresh runtime finishes steps 3 and 4

Guide

Reference