convopack

Framework-agnostic, provider-agnostic context-window packer for LLM chat history.

Why

LLM apps accumulate messages until they overflow the model's context window. Existing fixes are framework-locked (LangChain, LangGraph), provider-specific (Anthropic context-management beta), or designed for long-term semantic memory rather than turn-by-turn packing (mem0).

convopack is a small, focused library that takes a conversation history and a token budget and returns the largest tail that fits, while:

• preserving tool_use / tool_result pairs atomically,
• normalising message shapes across OpenAI, Anthropic, and Gemini,
• letting you plug in any tokenizer or summariser,
• staying async-friendly and zero-dependency at the core.

Install

pip install convopack                 # core only
pip install "convopack[tiktoken]"     # + OpenAI tokenizer
pip install "convopack[anthropic]"    # + Anthropic tokenizer
pip install "convopack[all]"          # everything

# or with uv
uv add convopack                       # core
uv add "convopack[all]"                # everything

The library ships a py.typed marker; mypy --strict and pyright both recognise its public types without further configuration.

Quickstart

from convopack import Packer, Recency

packer = Packer(
    budget=8000,
    tokenizer="tiktoken:gpt-4o",
    strategy=Recency(),
    pin=["system", "first_user"],
)

packed = packer.pack(messages)        # list[dict] in, list[dict] out

Strategies

Comparison

Example notebooks

Self-contained Jupyter notebooks. Every one runs end-to-end without an API key (they use the zero-dependency approx tokenizer and deterministic fakes). All four are executed in CI to stay in sync with the library.

Rendered with outputs at https://mrrobi.github.io/convopack/notebooks/01_quickstart/.

License

MIT