A fresh Show HN post dropped Friday introducing MACCHA (Multi-Agent Continuous Context Harness), an open-source project that tackles one of the most persistent frustrations with AI coding assistants: they forget everything between sessions. The framework establishes a file-based "brain" living directly in your home directory, giving any local agent—Antigravity, OpenCode, Claude Code, Cursor—a shared, evolving memory without requiring heavy background services or cloud dependencies.

The Core Problem MACCHA Is Solving

Modern AI coding agents start from absolute zero every time you open a new workspace. They forget your technical preferences (pnpm over npm, dark-mode CSSsans Tailwind), system constraints, and the hard-earned lessons from yesterday's debugging session. You repeat yourself constantly. MACCHA ends that loop by creating a structured memory hierarchy and automated maintenance loops that any agent reads automatically at startup. The architecture organizes context into seven distinct tiers with strict priority enforcement. TIER 0 (~/AGENTS.md) serves as the master bootstrap—the absolute first file any agent reads. TIER 1 handles project mandates and session rules. TIER 2 covers machine-level preferences like hardware profiles and global security constraints. The lower tiers handle situational context, live task state, auto-improvement feedback, and curated learned-lessons respectively.

Memanto: A Working Memory Engine With Teeth

The real intellectual firepower lives in the Memanto engine (brain/lib/memanto_engine.py), a 13-category temporal memory system that goes far beyond simple key-value storage. It implements exponential confidence decay—transient facts automatically lose weight over time while stable core preferences stay sharp. The semantic conflict detection scans proposed memory inserts against existing ones, catching and superseding contradictions before they compound. The hybrid search recaller combines cosine-similarity vector embeddings with Jaccard keyword overlap fallbacks, meaning retrieval works even when exact terminology drifts between sessions. This is specifically designed to be compatible with Gemini 3.0 Thought Signatures, giving you a path forward if you're already invested in Google's model ecosystem.

Supply Chain Safety Guards That Don't Get In Your Way

Security-conscious developers will appreciate MACCHA's approach to supply chain protection. The framework enforces a minimum-release-age=7d cooldown on all packages before they'll be considered for installation—catching zero-day exploits before they hit your dependency tree. A Human-in-the-Loop mandate blocks agents from modifying real-world capital or sending emails without explicit manual confirmation. Pre-install pnpm audit checks run automatically, printing Safety Report: GREEN/RED flags directly to the terminal. The secrets scanner detects leaked API keys, RSA private keys, and service JSONs before commits land in your repository—a sanity check that should ship with every AI coding tool by default.

Lightweight By Design—No Always-On Daemon Required

Unlike projects like Hermes Agent or OpenClaw that run as 24/7 background daemons consuming constant compute, MACCHA only consumes resources when you actively invoke an agent. The system runs flawlessly on constrained hardware: Chromebooks running Crostini, older laptops with limited RAM, free-tier cloud environments. You get the persistent memory and deep context of a sophisticated digital assistant while keeping full control and minimal resource overhead. The backup system uses AES-256 encryption to tar your personal zones (INFO/, PLAN/, BRAIN/) onto mounted cloud drives with a 4-backup rotation. The key stays local in ~/.config/maccha/backup.key—you keep that off-device or lose everything, which is exactly how it should work.

Key Takeaways

  • Seven-tier context architecture enforces strict priority—no memory bloat or conflicting rules
  • Memanto engine handles confidence decay and semantic conflict detection automatically
  • No background daemon required—agents only consume compute when invoked
  • Built-in supply chain safety: 7-day package cooldown, secrets scanning, pnpm audit integration
  • Human-in-the-loop by default for sensitive operations—no runaway autonomous actions
  • Cross-platform persistence: switch between Antigravity, OpenCode, Claude Code seamlessly

The Bottom Line

MACCHA fills a real gap for developers who want sophisticated agentic memory without the infrastructure overhead of always-on systems. If you've been burned by AI assistants that forget your preferences every session, this framework offers a clean, file-based solution that stays under your control. The seven-tier architecture and Memanto engine show genuine engineering thought—not just another wrapper around prompts. Worth watching as it matures.