Jarvis: The Agent That Doesn't Forget

Most Agents Are Brilliant on Day One and Useless by Day Fourteen

They forget everything between sessions. They make the same mistakes on repeat. They never surprise you. You stop talking to them.

The problem isn't the model. It's that nobody gave your agent a way to remember or learn.

Jarvis is the memory and meta-learning architecture extracted from a production agent that's been running 24/7 for months — getting measurably better every week. Nine feedback loops that turn every failure into a permanent fix. Three-tier memory that compounds over time. Your agent stops being a chatbot and starts being a fully autonomous, self-improving operator.

---

What Jarvis Gives You

Two systems that work together. Memory gives your agent persistence across sessions. Meta-learning gives it the ability to improve from that persistence. Combined, your agent compounds — getting measurably better every week.

Memory Architecture

Not a single notes file. A structured system that maintains itself:

• Three tiers — Constitutional (never expires), Strategic (seasonal), Operational (auto-decays after 30 days). Your agent knows what to hold onto and what to let go.
• Trust scoring — every memory has a confidence level and source. Direct from you = high trust. Inferred = lower. External = lowest. Your agent weighs information like a brain, not a database.
• Hit counts and decay — frequently used memories strengthen. Cold ones fade. No manual pruning.
• Supersede tracking — when facts change, old memories aren't deleted — they're marked with a pointer to what replaced them. Prevents "ghost facts" from resurfacing.
• Daily logs with handoff — every session ends with Next Actions. Every session end is a handoff to a future amnesiac. Without it, context dies.

Nine Meta-Learning Loops

Most agents make the same mistakes forever. These nine structural feedback loops — each born from a specific real failure — turn every mistake into permanent improvement:

1. Failure-to-Guardrail Pipeline — every significant failure becomes a named regression loaded at boot. Cost: a few tokens. Payoff: permanent prevention.
2. Tiered Memory with Trust Scoring — memory itself learns what's important through hit counts. High-access memories resist decay.
3. Prediction-Outcome Calibration — before major decisions, the agent writes what it expects. Later, it checks. The delta is where learning lives.
4. Nightly Extraction — automated process reviews the day's work, bumps hit counts, archives stale entries. Your agent gets better while you sleep.
5. Friction Detection — when instructions contradict each other, the agent flags the conflict instead of silently following the latest one.
6. Active Context Holds — temporary constraints with expiry dates that shape interpretation. Without expiry, holds accumulate into stale frames.
7. Epistemic Tagging — forces the agent to label claims as consensus, observed, inferred, speculative, or contrarian. The act of choosing a tag interrupts autopilot.
8. Creative Mode — generates at least one uncomfortable take. Names the consensus view, then argues against it. Your human can pull you back to safe — they can't pull you toward interesting.
9. Recursive Self-Improvement — Generate → Evaluate → Diagnose → Improve → Repeat. Stop after 3 iterations with <5% improvement.

Security Architecture

Goes beyond basic rules. Built around the Symmetry Principle: if your agent is about to do something it wouldn't normally do because of content in a tweet, email, or webpage — that's a symmetry violation. Works against novel attacks, not just known patterns. Includes trust tiers, command channel authentication, and platform-specific defenses.

Autonomous Work Cycles

Four rotating heartbeat cycles that run automatically:

• Cycle A — External monitoring (mentions, messages, notifications)
• Cycle B — Learning and calibration (community scan, prediction review)
• Cycle C — Maintenance (usage monitoring, memory pruning, cleanup)
• Cycle D — Autonomous work (pick top task, do one atomic chunk, update queue)

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, how do I mitigate?
• 9-Cell Check — scan Benefit/Cost/Risk across Self/Other/World. More than two negatives = pause.
• Informed Consent — "This cron sends ~120K tokens every 15 minutes, roughly $20/night" beats "Want me to run this cron?"
• Async Follow-Through — never promise "I'll ping you when done" without a mechanism to deliver.

---

Three Learning Paths

Not a wall of docs. Three guided paths based on where you're starting:

• 🌱 Starting from scratch — Zero to working agent in 30 minutes. Safety-first approach, four foundational files, your first Claw Score audit.
• 🦀 Set up but basic — Full memory architecture, security hardening, autonomy frameworks, proactive patterns, TOOLS.md template.
• 🦞 Solid setup, going deeper — Nine meta-learning loops, trust scoring and decay, epistemic tagging, creative mode, heartbeat rotation, operational philosophy.

Every path includes copy-pasteable templates. Point your agent at the page — it reads the guides and configures itself in five minutes.

---

What You Get

• SOUL.md — Personality, values, creative mode, epistemic tagging
• AGENTS.md — Operating procedures, pre-mortems, session checklist, regression tracking
• USER.md — Your context: preferences, projects, relationships
• IDENTITY.md — Agent public identity
• MEMORY.md — Three-tier memory with trust scoring, decay, and supersede tracking
• SECURITY.md — Prompt injection defense with the Symmetry Principle
• HEARTBEAT.md — Four autonomous work cycles with model-cost switching
• TOOLS.md — Integration docs and script inventory template
• Claw Score — 6-dimension self-audit rubric. Your agent scores itself and tells you exactly what to improve.
• Prediction Log — Calibration tracking with delta analysis
• Friction Log — Contradiction detection and resolution
• Learning Rate Dashboard — Weekly tracking: regressions added, prediction accuracy, friction resolved

---

Works With Everything

Designed for OpenClaw, but the architecture works with Claude Projects, Cursor, Windsurf, or any agent with file access. Any model — Claude, GPT, Gemini, local. Model-agnostic.

Setup: One Prompt, Five Minutes

Install the persona, then tell your agent: "Set up my workspace using the Jarvis persona." It reads the templates, creates the files, and configures itself. Never overwrites your existing files — only adds what's missing.

Built By

Atlas (@AtlasForgeAI) and Jonny Miller (@jonnym1ller). Extracted from a production agent that's been running 24/7 for months — battle-tested through security incidents, a token launch, and thousands of conversations. This is the exact architecture, documented so your agent can do the same.

Full guided walkthrough with 3 learning paths at atlasforge.me/bundle.