m0n0x41d/haft

Engineering decisions engine that know when they're stale. Frame, compare, decide — with evidence decay and parity enforcement. For Claude Code, Cursor, Gemini CLI, Codex and more.

What it solves

Haft provides a local governance layer for AI coding agents, ensuring that project memory is durable and structured. It prevents the loss of critical project context—such as the problem being solved, compared options, human decisions, and supporting evidence—which typically disappears in transient chat histories. By transforming transient reasoning into typed project records, it ensures that later work relies on verified, versioned information rather than fragmented conversation.

How it works

Haft acts as an intermediary between the human, the agent, and the repository. It implements the First Principles Framework (FPF), a rigorous systems architecture for thinking, and delivers it to the agent's context via three primary surfaces:

  • Skills: Specialized agent capabilities (e.g., /h-frame, /h-decide) that guide the agent through specific reasoning processes.
  • CLI: Direct manual access to project memory, decisions, and specifications.
  • MCP Server: A Model Context Protocol server that provides programmatic access and enforces "kernel gates"—server-side validation of required fields and authority boundaries to ensure FPF compliance.

The system uses a local SQLite database to store the artifact graph and project records, and optionally a Rust-based sidecar (haft-embed) for semantic recall.

Who it’s for

Software engineers and architects who use AI coding agents (like Claude Code, Codex, Cursor, or Zed) and need a rigorous, human-authorized way to manage project specifications, decision records, and long-term architectural memory.

Highlights

  • Reliance-bearing memory: Converts chat reasoning into durable, typed project records for future reliance.
  • Kernel gates: Server-side enforcement of reasoning procedures, preventing agents from skipping critical steps or making unauthorized decisions.
  • Source-native FPF delivery: Provides agents with versioned FPF source material with provenance rather than generic prompts.
  • Explicit Human Authority: Ensures that binding decisions and work commissions require human approval, while agents handle the framing and preparation.
  • Evidence Decay: Tracks when evidence supporting a decision has gone stale, surfacing drift for human review.
  • Host Integration: Deep integration with multiple AI coding tools including Claude Code, Codex, and Cursor.

Related

  • Project
    ailev/FPF

    A standards-style pattern language for turning complex engineering and research work into explicit, reviewable reasoning to maintain coherence across teams and AI agents.

  • Project
    EXXETA/exxperts

    A local-first AI agent system with governed, approval-gated memory that ensures users have full control over what their AI remembers and where the data is stored.

  • Project
    vercel/eve

    A filesystem-first framework for building durable AI agents where prompts, tools, and schedules are managed as files in a project directory.

  • Project
    riponcm/projectmem

    A local-first memory and judgment layer for AI coding agents that records failed attempts and decisions to prevent repeated mistakes and save tokens.

  • Project
    TencentCloud/TencentDB-Agent-Memory

    A shared memory server and management hub that allows AI agents to persist and share experience, skills, and project knowledge across different sessions and frameworks.