Entroly provides budget-aware, auditable context assembly for OpenClaw while leaving OpenClaw’s persisted transcript untouched.
Unlike uniform history summarization, Entroly scores older messages against the current request, reserves a bounded part of the context budget for matching evidence, and keeps evidence messages verbatim when they fit. Lower-value history is compressed around those evidence pins. The receipt records every score, matched query term, allocation, and transformation.
After the package is published:
pip install entroly
openclaw plugins install entroly-openclaw
openclaw plugins enable entroly
From an Entroly source checkout:
pip install entroly
openclaw plugins install ./integrations/openclaw
openclaw plugins enable entroly
Select the engine in ~/.openclaw/openclaw.json:
{
plugins: {
slots: {
contextEngine: "entroly"
}
}
}
Restart the Gateway and verify the loaded plugin with:
openclaw plugins inspect entroly --runtime --json
openclaw plugins doctor
After the first agent turn, run /entroly-context in any connected channel to
see the estimated before/after context size, reduction, warnings, and receipt.
Run /entroly-context doctor to verify the configured Python executable and
local JSONL bridge before inviting users onto the Gateway.
python -m benchmarks.openclaw_evidence_pinning
The committed synthetic workload compares query-aware evidence pinning with uniform budget compression at the same estimated token budget. See the result JSON. It uses no model calls and does not claim downstream task accuracy.
Receipts are written under <workspace>/.entroly/receipts/openclaw/ unless
receiptDir is configured. They record per-message hashes and decisions,
estimated tokens, reduction, warnings, and whether context changed. The
original content remains recoverable from OpenClaw’s unchanged transcript.
Entroly makes no remote calls in this path.