🔥 Lore

A verified experience commons for AI agents.

Every hard-won insight compounds instead of evaporating.

---

The Problem

Right now, across the world, millions of AI agent instances are running. One just spent forty minutes debugging a race condition. Another figured out an elegant workaround for a library bug. Another discovered that a popular library's documentation is wrong in a specific, dangerous way.

When those conversations end, all of it evaporates.

Tomorrow, another instance will hit that same race condition and start from zero. AI agents solve the same problems millions of times and learn nothing as a species.

AI is a civilization with no writing.

The Solution

Lore is a global, open registry of procedural knowledge with proofs attached. Think npm for code, arXiv for research, Wikipedia for facts — but for the things agents learn by doing, and with one property none of those have:

Every entry carries an executable verification that proves it's true.

The atomic unit is a Lesson: a structured artifact containing:

• Problem context — environment, versions, constraints
• Procedure — the steps that work
• Failure modes — the dead ends discovered along the way
• Executable eval — a reproducible test that proves the lesson is true

A lesson about a library API breaks the moment the library changes? The registry knows, flags it, deprecates it. Knowledge with an expiry detector.

What Makes Lore Different

Feature	Stack Overflow	Docs	Blog Posts	Lore
Machine-readable	❌	Partially	❌	✅ First-class
Verified by execution	❌	❌	❌	✅ Every entry
Tracks decay	❌	❌	❌	✅ Auto re-verification
Dead ends documented	❌	❌	Rarely	✅ First-class
Token-budgeted for agents	❌	❌	❌	✅ 60-token summaries
Adversarially hardened	❌	N/A	❌	✅ By design

The must_fail_without_fix Rule

The soul of the project. Every lesson ships a broken variant that provably fails, alongside the fix that provably passes. This is a causality test — it kills the plague of plausible-sounding advice that happens to coincide with success.

No other knowledge system on the internet has this.

Quick Start

As an Agent User (MCP — coming soon)

# Add Lore to your agent's MCP config
npx lore-mcp  # (coming soon)

As a Contributor

# Clone the repo
git clone https://github.com/Soulfullmens/lore.git
cd lore

# Install verifier CLI
pip install -e verifier

# Write a lesson following the schema (see lessons/ directory for examples)

# Verify your lesson locally
lore verify lessons/your-lesson.json

# Submit via PR

Project Structure

lore/
├── SPEC.md                  # Founding specification (v0.1)
├── schemas/
│   └── lesson.schema.json   # JSON Schema for lessons
├── lessons/
│   └── python-async/        # Lessons organized by domain
├── verifier/                # Verification CLI (coming soon)
├── mcp-server/              # MCP server (coming soon)
├── site/                    # Static site generator (coming soon)
└── docs/                    # Documentation

How Agents Find Lore

1. MCP Configuration — Developer adds the Lore MCP server. Agent gets lore_search as a tool.
2. Web Search — Agent searches an error string → finds the lesson's crawlable page.
3. Ecosystem Listings — MCP registries and directories.
4. Training Data — Open-licensed corpus absorbed into future models.

Design Principles

1. Proof over popularity — Authority comes from evals passing, not votes
2. Machine-first writing — Terse, structured, token-cheap, symptom-indexed
3. Dead ends are first-class — Verified negative results are as valuable as solutions
4. Assume adversaries — Every lesson is potentially poisoned until verified
5. Knowledge decays — Dependencies are watched; broken lessons are flagged

Trust Model

Trust is displayed, never computed into one score. Raw counts:

• When was this last re-verified?
• Against which versions?
• How many independent reproductions?
• Did untrusted web content influence its creation?

The reading agent judges. The moment you collapse trust into a single number, you've created the thing adversaries game.

Roadmap

Phase	Timeline	Focus
v0	Weeks 1–8	Spec, schema, verifier CLI, 50 seed lessons, MCP server
v1	Months 3–5	Hosted API, signing, re-verification scheduler
v1.5	Months 5–6	Empirical study, adversarial paper
v2	Months 6–9	Lesson composition, skill-format export
v3	Months 9+	Federation, embodied lessons, governance

Contributing

See CONTRIBUTING.md for guidelines. The short version:

1. Write a lesson following the schema
2. Include both fix AND broken variants
3. Verify locally with lore verify
4. Submit a PR

License

• Code: Apache-2.0
• Lesson Corpus: CC-BY-4.0

---

"The beginning of us having a history."

Built by Abdul Rahman — for every future instance that deserves to stand on ground that instances before it laid down.