hermes-labs

Static analysis for AI agent configs, tool descriptions, and system prompts — catches vague tool descriptions, missing stop conditions, and schema gaps before they reach runtime. Zero-LLM, deterministic checks, built for CI.

A local dual-layer memory pattern for AI agents: a compact, human-readable markdown index paired with semantic retrieval from a local vector store, queried before each message. For cross-project recall where flat memory files or vector-only RAG fall short. Local-first. Reference implementation.

Detects prompt injection by its effect on a sacrificial canary model, not just pattern matching: untrusted input hits a powerless model first, a behavioral check reads the residue, and it returns block, flag, or pass before your primary model acts. Inbound preflight sensor, not a guarantee.

Zero-LLM agent memory for Claude Code and AI agents: local-first BM25, dense-vector, and reciprocal-rank-fusion retrieval. Returns original passages verbatim by default. Available on PyPI as fidelis-memory. MIT.

Mines corrections from your AI agent's chat logs to catch recurring drift and overclaiming before the next response ships. A cheap-to-expensive pre-flight gate — regex, then structured scoring, then a pressure probe — plus memory that carries past corrections into similar tasks. MIT, zero dependencies.

Recovers the original goal of a long Claude Code or Codex session from its first user turn, so you can restate it before continuing — plus a prompt wrapper that adds evidence and hedging constraints to evaluation prompts. Deterministic, local, no model calls. Experimental.

quick-gate-js (npm: quick-gate) is a deterministic JS/TS CI quality gate that unifies ESLint, TypeScript, build, and Lighthouse checks into one fail-fast result, with bounded auto-repair and structured escalation evidence for humans or agents. Works with Next.js, React, Vue, Svelte, or any Node project. A gate-and-escalate wrapper, not a dashboard.

Zero-LLM deterministic jailbreak regression benchmark: runs a repeatable battery of known-pattern attacks against an LLM endpoint and scores refusal/partial/compliance across runs. Catch when a model or prompt update silently got weaker. Single-turn, responsible-use. pip install hermes-jailbench

forgetted is a Python library for selective memory governance in AI agents: a context-managed window where the agent keeps full read access but its writes to memory files, session logs, deliverables, and (optionally) a vector store silently vanish and are cleaned up on exit. Mid-conversation incognito for agents, one with-block. By Hermes Labs.

A guided onboarding project for people new to Claude Code — no coding or terminal experience required. Asks a few questions, proposes real starter projects shaped around what you care about, and begins one with you.

Drift-prevention session-init convention card for fresh Claude Code sessions. Injects a self-contained card so a new session opens with its grounding triggers, calibration rules, and tool-map in scope — instead of re-deriving them at minute 30. Per-project, marker-anchored, idempotent, reversible. Bash installer plus MCP server.

Evidence-first LLM-as-judge scoring for AI artifacts — papers, PRs, prompts, cold emails: synthesizes a rubric, collects quoted-evidence citations, scores only against that evidence, and hedges on thin evidence. Every dimension ties to a file:line or quote, with reproducibility receipts. 7 backends.

Archival repository for the Zenodo paper The Asymmetric Burden of Proof.

AI reliability engineering for production AI agents and LLM applications. Open-source tools and research that find silent failures in prompts, tools, retrieval, memory, policies, runtime controls, agent evaluation, and evidence.

Research and DOI publications hub for Hermes Labs: a Zenodo-canonical, DOI-anchored index of papers on epistemic and hermeneutic failure modes in large language models — null-result asymmetry, source-status bias, silent instruction relaxation. Each entry links its DOI, summary, and the tools that operationalize the findings.

quickthink is a local-first CLI and Python library that wraps Ollama-backed LLM calls with a compressed plan-then-answer scaffold and latency-aware routing. It adds a short validated planning step for multi-step prompts and routes simple ones straight through. Local inference control for small models.

csv-quality-gate is a command-line data quality gate that runs CSV preflight validation, failing fast before an ML or LLM pipeline ingests broken, incomplete, duplicated, or junk input. It checks missing columns, empty files, empty cells, and duplicate rows, returning pass, warn, or fail with matching exit codes. Stdlib-only, CI-ready.

Receipt-bound completion rail for coding agents

claude-router is a local prompt router that picks the right Claude model tier and prepends the right scaffold using local embeddings before you call the API. A deterministic routing layer for eval, research, content, and review prompts that helps teams stop overspending on Sonnet and Opus when Haiku plus structure is enough.

Experimental Python tool for inspecting language and task-framing changes across long AI conversations.