bioma-micro

B.I.O.M.A. Micro-Kernel — a lean efficiency & resilience core for LLM infrastructure, written in Rust, exposed to Python via PyO3.

Exactly two proven primitives, nothing else:

• Context apoptosis — autonomous history dehydration. Long agent sessions accumulate low-value ballast (verbose tool logs, stale turns). The kernel assigns each message a metabolic weight by class, applies half-life decay, and purges dead weight before the payload hits the API. Microsecond-scale, GIL-released, allocation-light.
• Hormonal bus — lock-free in-memory signal injection (~2M signals/s, ~5 µs).

Plus saturation_scan(), an O(n) repetition detector for cognitive-DDoS / flood inputs.

Install

pip install bioma-micro

Prebuilt wheel for Windows (Python ≥ 3.8, abi3). Other platforms build from the sdist (requires a Rust toolchain: pip will invoke maturin automatically).

Quick start

import bioma_micro as bm

messages = [
    ("You are a precise operations copilot.", bm.SYSTEM),   # never purged
    ("FACT: the deploy freeze ends on 2026-07-18.", bm.FACT),  # never purged
    ("[tool log] 40 KB of verbose audit output ...", bm.TOOL),  # prime target
    ("Round 1: any anomaly?", bm.USER),
    ("Nothing above baseline.", bm.ASSISTANT),
    # ... hundreds of turns later ...
    ("What is the freeze end date?", bm.USER),
]

result = bm.dehydrate(messages, half_life=6.0, safe_threshold=0.35)
print(result["reduction"])          # fraction of input tokens purged (0..1)
print(result["kernel_latency_us"])  # pure-Rust decision pass, microseconds
prompt = "\n".join(result["kept"])  # dispatch this instead of the full history

Flood detection:

bm.saturation_scan(suspicious_text)  # ~1.0 = repetitive flood, ~0.0 = natural text

The honest contract

Measured, auditable behavior (full methodology and raw data in the framework repo):

• Values tagged FACT and recent turns survive; answer quality stays at parity with the full-context baseline (verified with objective probes against real online models).
• Durable info buried in old untagged turns is purged by design — tag it FACT.
• Against an already-lean agent the kernel is a no-op: it never helps, never hurts.

This is a safety net for agents that do not manage their own context — not a universal "−X%" claim.

API surface

Cache-aware dehydration & thinking budgets (1.1.0)

Two cost phases, two levers:

• Input (prefill): apoptosis is deletion-only and order-preserving, so a durable prefix stays byte-identical. With provider prompt caching, pass stable_prefix to guarantee it, and ask consolidation_gain() when purging cached ballast actually pays off (cache reads at 0.1× make stale prefixes cheap — break-even is often 15+ calls away; the kernel does this arithmetic for you).
• Reasoning (decode): thinking tokens are output-priced (~5×) and sequential. effort_gauge() scores each request with cheap lexical signals and returns the budget to forward as budget_tokens (Anthropic) / reasoning_effort (OpenAI) — trivial turns ("yes", "continue") stop paying for a full thinking budget.

License

FSL-1.1-MIT — Functional Source License 1.1 with MIT future license. Free for internal use, non-commercial use, and professional services; converts to MIT two years after each release.