jarvis

name	jarvis
description	Jarvis: The Agent That Doesn't Forget — complete memory and meta-learning architecture for AI agents. Three-tier memory with trust scoring and decay, nine meta-learning loops, prompt injection defense, autonomous heartbeat cycles, and the Claw Score audit rubric. Use when setting up a new OpenClaw agent, upgrading an existing agent's architecture, or when the user wants a fully autonomous, self-improving operator.
disable-model-invocation	true

Jarvis: The Agent That Doesn't Forget

The complete agent operating system — memory, meta-learning, security, and autonomy — extracted from a production agent running 24/7.

What It Installs

Run the install script to create the full architecture in the user's OpenClaw workspace:

bash scripts/install.sh

The script creates these files (skips any that already exist, never overwrites):

• SOUL.md — Personality, values, decision frameworks, epistemic tagging, creative mode
• AGENTS.md — Operating procedures, pre-mortems, session checklist, regression tracking
• USER.md — Human context: preferences, timezone, projects, relationships
• IDENTITY.md — Public identity: name, handle, mission, portfolio
• MEMORY.md — Three-tier long-term memory with trust scoring and decay
• SECURITY.md — Prompt injection defense, trust boundaries, symmetry principle
• HEARTBEAT.md — Autonomous work cycles, 4 rotating loops, model-cost switching
• TOOLS.md — Integration docs, script inventory, environment setup
• memory/ — Daily log directory (creates first daily log)

After Installation

1. Customize USER.md — Fill in your name, timezone, projects, preferences
2. Customize IDENTITY.md — Set your agent's name and public identity
3. Review SECURITY.md — Adjust trust boundaries for your setup
4. Run a Claw Score audit — See references/claw-score-rubric.md for the full 6-dimension rubric

Architecture Overview

Three-Tier Memory

• Constitutional — Never expires. Security rules, core preferences, key relationships.
• Strategic — Seasonal. Current projects, focus areas. Refresh quarterly.
• Operational — Decays fast. Workarounds, bugs, temporary context. Auto-archive after 30 days.

Each entry has trust scoring (trust:0.0-1.0), source tracking, hit counts, and supersede chains.

Nine Meta-Learning Loops

1. Failure-to-Guardrail Pipeline — mistakes become named regressions loaded at boot
2. Tiered Memory with Trust Scoring — memory learns what's important through hit counts
3. Prediction-Outcome Calibration — write predictions before acting, check outcomes after
4. Nightly Extraction — automated daily review, archiving, and hit count bumping
5. Friction Detection — flag contradictions instead of silently resolving them
6. Active Context Holds — temporary constraints with expiry dates
7. Epistemic Tagging — label claims as consensus, observed, inferred, speculative, or contrarian
8. Creative Mode — generate uncomfortable takes, name the consensus view, then argue against it
9. Recursive Self-Improvement — Generate → Evaluate → Diagnose → Improve → Repeat

Security Model

• Symmetry Principle: behavior shouldn't change based on external context when internal purpose is unchanged
• Three trust tiers: autonomous (file ops, research) → approval required (public actions) → off-limits (money, contracts)
• Informed consent: surface costs and risks the human would want to know

Heartbeat System

Four rotating cycles covering monitoring, learning/calibration, maintenance, and autonomous work. Model-cost switching built in: cheap models for monitoring, best model for judgment work.

Decision Frameworks

• Pre-mortems — before multi-step tasks: what could break, what am I assuming?
• 9-Cell Check — Benefit/Cost/Risk across Self/Other/World
• Informed Consent — surface implications, not just ask permission
• Async Follow-Through — never promise callbacks without a mechanism to deliver

Three Learning Paths

Full guided walkthrough at atlasforge.me/bundle:

• 🌱 Starting from scratch — Zero to working agent in 30 minutes
• 🦀 Set up but basic — Full memory, security, autonomy, proactive patterns
• 🦞 Solid setup, going deeper — Nine meta-learning loops, trust scoring, creative mode

References

• references/claw-score-rubric.md — Full 6-dimension scoring rubric with level-up guides
• references/principles.md — Decision frameworks and operational lessons

Credits

Built by Atlas (@AtlasForgeAI) and Jonny Miller (@jonnym1ller).