bioma-langchain 0.1.0

Context apoptosis for LangChain: µs-scale, class-aware history pruning (B.I.O.M.A. kernel) as a drop-in Runnable.

bioma-langchain

Context apoptosis for LangChain: prune stale, low-value history with the B.I.O.M.A. Rust kernel (~1 µs decision) before messages reach the model.

Custom LangChain agents that resend a growing history are exactly where the kernel measured −84% input tokens (and up to −97.6% on bloated sessions) with answer quality at parity — see the repo's honest benchmark report, including where it does not help.

Install

pip install bioma-langchain

Use

from bioma_langchain import BiomaDehydrator
from langchain_core.messages import HumanMessage

dehydrator = BiomaDehydrator()      # tool-calling default: safe_threshold=0.2

# LCEL — drop it right before the model:
chain = dehydrator | llm

# or call it directly:
lean_history = dehydrator.invoke(messages)
print(dehydrator.last_audit)        # tokens_before/after, reduction, µs

Durable values must be marked to survive forever (the usage contract):

HumanMessage("The rollback token is RBK-5521.",
             additional_kwargs={"bioma": "fact"})

Class mapping

LangChain message	Kernel class	Behavior
SystemMessage	SYSTEM	never purged
additional_kwargs={"bioma": "fact"}	FACT	never purged
HumanMessage / AIMessage	USER/ASSISTANT	decay by recency
ToolMessage, tool-call-only AIMessage	TOOL	dehydrates first

Declared limit: an old, untagged value decays below the threshold and is purged by design — tag durable info as fact.