Continuum

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

Agent workflows burn money recomputing what they already know: the same system prompt tokenized ten thousand times, the same subtask answered again, an hour-long run lost to one crash at step 19. Continuum is a C++ execution engine that treats LLM calls and tensor ops as operators in one dataflow graph — redundant work is cached at the runtime level, and a running workflow can be checkpointed to bytes, resumed in another process, or forked from any past step.

flowchart LR
    P([prompt]) --> M{memo}
    M -->|exact hit · 0 ms| R([result])
    M --> S{semantic}
    S -->|paraphrase hit · 0 ms| R
    S --> T{trie prefix KV}
    T -->|"shared prefix · ~99% fewer tokens sent"| B
    T --> L{layer KV}
    L -->|warm decode state| B[backend call]
    B --> R

Loading

• 92.5% token reduction on a mixed 20-step agent workload against live Azure OpenAI
• Zero-cost exact repeats — memoized calls skip the backend entirely
• Durable execution — checkpoint / crash / resume / time-travel fork, deterministic replay
• One graph for tokens and tensors — Azure, OpenAI, Anthropic, vLLM, libtorch, and MLX behind one IR

What You Can Build With It

• Agents that survive anything — deploys, crashes, spot-instance eviction: checkpoint mid-run, resume on another machine with the KV cache still warm.
• Cheaper agent fleets — hundreds of sessions sharing one system prompt send it once; the trie prefix cache serves the rest.
• Eval and CI loops — re-running near-identical prompt suites hits the memo and prefix tiers instead of your API budget.
• Time-travel debugging — rewind a finished run, edit step 7, replay the alternate timeline; completed steps come from the checkpoint, never recomputed.
• Hybrid pipelines — hosted LLM calls and local tensor ops (libtorch, MLX) as operators in the same scheduled graph.

How It Fits With What You Already Use

Continuum sits below your framework, not beside it — LangChain/LangGraph code, raw SDK calls, or plain Python all route through the same runtime.

You may already use	What it gives you	What Continuum adds
Provider prompt caching (OpenAI / Anthropic)	Prefix discounts inside one provider, TTL-bound	Provider-agnostic reuse, plus memo/semantic tiers and cache state you own
GPTCache / LangChain cache	App-layer response cache	Runtime-level reuse with defined invalidation semantics — tool calls never served stale
LangGraph checkpointing / Temporal	Workflow state persistence and retries	Checkpoints that carry the KV cache too — resume warm, deterministic replay, fork any past step
vLLM prefix caching	KV reuse on GPUs you operate	The same idea extended across hosted APIs, portable inside checkpoints

Quick Start

python -m pip install continuum-ai

Kill an agent mid-run and finish it in a different process:

from continuum._native import DurableAgent

agent = DurableAgent()
agent.begin(["research the topic", "draft the report", "publish it"])
ckpt = agent.run_until_step(1)        # bytes: graph + every value + KV cache state

# ... process dies here ...

revived = DurableAgent()              # brand-new runtime
outputs = revived.resume_from(ckpt)   # completes steps 3+ without redoing 1-2

Rewind a finished run, edit one step, and replay the alternate timeline:

forked = DurableAgent.fork(ckpt, node_id, "write a haiku instead")
alternate = DurableAgent().resume_from(forked)

See every reuse tier fire in one deterministic run:

PYTHONPATH=python python examples/05_continuum_reuse_stack.py   # --trace for per-tier firing
PYTHONPATH=python python examples/06_durable_agent.py           # checkpoint / crash / resume
PYTHONPATH=python python examples/07_time_travel_fork.py        # rewind, edit, replay

Measured Results

Isolated per-tier benchmarks against a live Azure OpenAI backend (gpt-5-mini), one reuse mechanism enabled at a time (benchmarks/v11/):

Mechanism	Workload	Result
Trie prefix KV cache	10 calls, 3,000-char shared prefix	~99% token reduction (9/9 hits, ~30 tokens sent per call)
Memo table	5 exact-repeat tool calls	5/5 backend calls skipped, 0 ms
Mixed 20-step agent workflow	prefix + repeats + paraphrases + cold queries	92.5% token reduction, 4/20 backend calls eliminated
Cross-session cold start	persist cache metadata, restart, reload	≥80% hit rate on first warm run
No-reuse worst case	4 unrelated queries	~0.5% overhead-free passthrough, no errors

Latency on prefix hits drops ~31% (5.4 s → 3.7 s median) — the API round-trip dominates once 99% of prompt tokens are skipped; token cost is where reuse pays. The bundled n-gram embedding provider is a placeholder: semantic-tier results require a real embedding model and are excluded from the headline numbers.

Raw data, per-experiment reports, and the scripts that produced every number live in benchmarks/v11/; plots in plots/v11/. Deterministic, CI-checked versions of every mechanism run offline via the FakeLLM backend (examples/05–07, tests/python/).

Why a Runtime, Not a Wrapper

Caching bolted onto an SDK can't know what is safe to reuse. Continuum sits below the program, where reuse has defined semantics:

• Correct invalidation — tool calls are never served from cache (side-effecting), memoized results are version-bumped on resume, and cached KV state is reused only when its tokens are verifiably a prefix of the query.
• Policy-gated — every tier respects a per-session ReusePolicy (always / never / prefix-length threshold): one switch, no stale reads.
• Portable state — backends that can export their state handles carry the KV cache inside the checkpoint, so a resumed process starts warm, not cold.
• Capability dispatch — backends declare tensor/token/cache capabilities; the scheduler routes each node, converting tensors across backends explicitly.

flowchart TB
    subgraph app["your code"]
        A[LangChain / SDK calls / plain Python]
    end
    subgraph rt["Continuum runtime (C++)"]
        IR[dataflow IR] --> SCHED[capability-aware scheduler]
        SCHED --> REUSE[five-tier reuse stack]
        SCHED --> CKPT[(checkpoints<br/>graph + values + KV state)]
    end
    subgraph backends["backends"]
        B1[Azure / OpenAI / Anthropic]
        B2[vLLM]
        B3[libtorch / MLX]
    end
    A --> IR
    REUSE --> B1 & B2 & B3

Loading

What Is Implemented

• C++ execution engine with IR interpreter and serializable checkpoints
• Five-tier reuse stack: trie prefix KV cache, memo table, semantic cache, layer KV warm-start, memory graph recall
• Durable execution: checkpoint a running workflow to bytes, resume in a fresh process (KV cache included), or fork from a past step with an edited value
• Session API with per-tier reuse policies and cross-session cache persistence
• Backends: Azure OpenAI, OpenAI, Anthropic, vLLM shim, libtorch, MLX, deterministic FakeLLM for CI

Current Status

• v1 release hardening in progress
• CIR schema lock with serialization conformance (schema/cir.fbs)
• Linux and macOS CI matrix with coverage gates and fuzz workflow
• PyPI packaging under continuum-ai (import path remains continuum)

Documentation

• Python API docs: https://ct.rithul.dev/python/
• C++ API docs: https://ct.rithul.dev/cpp/

Build docs locally:

# Python docs
python -m venv .venv-docs
. .venv-docs/bin/activate
pip install sphinx furo breathe
PYTHONPATH=python sphinx-build -b html docs/api/python docs/api/python/_build

# C++ docs
doxygen Doxyfile

Local outputs:

• docs/api/python/_build/index.html
• docs/api/cpp/html/index.html

Community

• Contributing guide: CONTRIBUTING.md
• Code of Conduct: CODE_OF_CONDUCT.md
• Security policy: SECURITY.md
• Support guide: SUPPORT.md
• Governance: GOVERNANCE.md

Quick contributor setup:

pip install pre-commit
pre-commit install
pre-commit run --all-files
pytest

Citation

If Continuum helps your work, cite it as:

@software{continuum2026,
  title        = {Continuum: Unified Runtime for Token and Tensor Programs},
  author       = {Kamesh, Rithul and Contributors},
  year         = {2026},
  url          = {https://github.com/rithulkamesh/continuum},
  version      = {1.0.0}
}