Entroly for OpenClaw
A first-class OpenClaw context engine with local bridge processing, Context Receipts, provider-independent budgeting, and optional proof-guided exact-message recovery.
OpenClaw context engine guide →Entroly is the open-source Context Assurance layer for AI agents. It helps each host select evidence under a budget, record what was included or omitted, preserve exact originals behind content-addressed handles, and verify claims without replacing the host runtime or model provider.
A first-class OpenClaw context engine with local bridge processing, Context Receipts, provider-independent budgeting, and optional proof-guided exact-message recovery.
OpenClaw context engine guide →A Hermes ContextEngine implementation with request-only selection, model-window updates, usage observation, status reporting, and hash-only full-conversation recovery.
Hermes context engine guide →A local MCP setup plus a compaction hook that protects paths, symbols, commands, errors, verification status, rejected hypotheses, and every exact recovery handle.
OpenCode Context Assurance guide →Across integrations, discovery and recovery stay separate. Ranking may identify useful evidence, but a known ccr:<24-hex> handle is resolved through a deterministic, hash-only lookup that returns the complete stored original. A missing historical handle does not silently resolve to a newer source revision.
entroly_retrieve({"hash": "ccr:0123456789abcdef01234567"})
No. The host remains responsible for its agent loop, transcript, permissions, provider authentication, tool execution, retries, and user experience.
No universal guarantee is claimed. Entroly exposes local context measurements, provider-bound evidence where observable, receipts, recovery, and verification so teams can evaluate their own workload. See the AI cost optimization guide.
Yes while the referenced CCR entry remains retained. Exact recovery uses the content handle, not fuzzy search or a generated summary.
The context-selection, receipt, verification, and CCR paths are local-first. A host may still send the selected prompt to the model provider configured by the user.