Saphene

Saphene is CI for AI agent memory — "Clarity Through Evidence." Coding agents summarize decisions, teams write specifications, and projects accumulate confident statements. Then the code changes, and nobody knows which memories are still true. Existing tools scan code for known bad patterns; Saphene verifies whether your documents, requirements, and agent memory still match the code. How it works: 1. Claims in — docs, specs, SRS requirements, and agent memory become structured claims with literal evidence hints. 2. Evidence out — a deterministic search locates (or fails to locate) each claim's evidence in the repository. Every verdict (matched / partial / mismatch / unfound) carries a file:line receipt with a verbatim anchor quote. 3. Ledger — verdicts persist with stable claim IDs. The next run diffs against the last: new, resolved, relocated, or drifted. 4. Enforce — the GitHub Actions PR gate fails when a verified claim regresses; the ledger diff is posted as a sticky PR comment. Saphene also checks documents against each other for contradictions, handles negative and planned claims honestly (absence inside a declared scope, future claims verified against the plan), supports per-repo alias/disambiguation tables, and runs a deterministic rule scanner plus SARIF ingestion so Slither/Semgrep/Snyk findings join the same ledger. It integrates where developers already work: a local-first CLI (saphene), an MCP server for coding agents (OpenCode, Cursor, Cline, Copilot), and GitHub Actions. Every model path is optional — pass --no-model for fully offline, deterministic verification with no API key. GitHub: https://github.com/Jamespino20/Saphene Site: https://metahackathon-saphene.vercel.app