v0.1.0 — bounded-context memory for long-horizon agents

What this is

A tiered memory controller for long-horizon LLM agents. It bounds the volume of
the mutable context (tool output, scratch work) so the invariant context (system
prompt, standing goal) keeps a fixed share of the softmax denominator regardless
of how many turns the run goes on for.

Code is Apache-2.0. Prose and paper are CC BY-NC-ND 4.0.

What the bundled simulation measures

Deterministic, 3 scenarios × 40 turns, no model inference, no API calls:

arm	active tokens	invariant attention share	CSR
append-only history	5,099	0.0425	0.0059
sliding window	3,340	0.0553	0.0077
generational context	129	0.0757	0.2000

Reproduce in about a second:

pip install -e ".[test]"
python ccb_bench.py

What this does not claim

• No live-model evaluation. That is the obvious next step and it is not done.
• The softmax inundation bound is elementary — Cauchy–Schwarz applied to the
  softmax — and it is vacuous under realistic projection norms. §8 of the paper
  says so in plain text.
• The defensible result is Theorem 2: a worst-case allocation guarantee that
  holds when the volume bound is actually maintained as a post-condition.

Attached

File	What it is
The_Softmax_Inundation_Theorem.pdf	Compiled paper, 8 pages
ccb_results.csv	Per-turn telemetry for all 3 scenarios
CCB_BENCHMARK_REPORT.md	Generated report with the derived table