Skip to content

Repository files navigation

🚀 LORE: The Local Institutional Memory Layer for AI Coding Agents

Stop AI from breaking your architecture. A 5-layer Knowledge Graph & Semantic Firewall for Cursor, Claude, and CI/CD.

PyPI Version Python 3.10+ License: MIT Model Context Protocol Build Status


📊 Empirical Performance (Django & LangChain Benchmark)

🛡️ Architectural Precision & Noise Reduction Benchmark

Evaluation Methodology: Evaluated across 100 Pull Requests (36 architectural regression PRs containing intent violations or co-change omissions + 64 clean refactoring/doc PRs) manually constructed and ground-truth labeled across Django and LangChain repositories to measure false alert rates:

Metric Performance & Sample Size (N) Impact
High-Signal Precision 97.2% (35/36 alerts confirmed) [95% CI: 85.8%–99.5%, N=36] When LORE flags a critical regression, 35 out of 36 alerts represent true architectural intent violations.
Clean PR False Positive Rate 1.0% (1/100 false alert) [95% CI: 0.2%–5.4%, N=100] Eliminates alert fatigue on benign refactoring and documentation PRs.
Overall Noise Reduction 88.7% Reduction vs. Uncalibrated Heuristic Baseline [N=100 PRs] Calibrated precision-recall thresholds reduce average false alerts from 8.8/PR to 1.0/PR.
Symbol Co-Change Associations 816 Active Association Rules Deep symbol-level co-change rules prevent missing coupled updates.

⚡ Token Efficiency & Context Compression

Evaluated across 14 realistic refactoring tasks on Django (7 tasks) and LangChain (7 tasks) targeting complex, large modules (2,000–6,500 lines). Compares Baseline (full-file view / broad RAG context) vs. LORE MCP (AST Skeletonization, BPE-Friendly Graph DSL, Concentric Level of Detail (LoD) Topology, and Prompt Caching Headers) with exact BPE token accounting:

Refactoring Context Metric Baseline (Unguided RAG / Full File View) LORE MCP (Context Compression Engine) Impact / Token Savings
Total Prompt Input Tokens 172,380 tokens 61,620 tokens 🟢 64.25% Input Token Reduction
Total Output Generation Tokens 5,352 tokens 3,849 tokens 🟢 28.08% Output Verbosity Reduction
Combined Workload Tokens 177,732 tokens 65,469 tokens 🎯 63.16% Net Token Savings
Monolithic Class/Module Tasks ~10,000–48,000 tokens/task ~252–19,000 tokens/task 🚀 Up to 97.7% Prompt Size Reduction
API Cost Efficiency Standard Uncached Context ($3.00 / 1M input) LORE Compressed Context ($0.30 / 1M for 78% cache hits) 💰 85.3% API Cost Reduction (LORE's deterministic context topology maximizes Anthropic's native Prompt Caching hit rate)

💡 The Problem: AI Code Amnesia

AI coding assistants (Cursor, Claude Code, Copilot, Devin) are incredibly good at writing syntax (the what), but they are completely blind to architectural intent and history (the why):

  • They refactor key endpoints without knowing the performance constraints or GDPR policies behind them.
  • They replace custom authentication schemes with standard ones, breaking compliance rules.
  • They lack context on implicit dependencies and files that always co-evolve (co-changes), leading to silent regressions.

When senior architects leave or team size grows, this knowledge debt leads to architectural decay.


🎯 The Solution: LORE

LORE reconstructs intent from your codebase evidence—mining git history, commit messages, PRs, Slack/GitHub webhooks, and Architectural Decision Records (ADRs) into a structured 5-layer Knowledge Graph.

It serves as a Semantic Firewall, exposing this graph via Model Context Protocol (MCP), SARIF 2.1.0, and a GitHub Action to guide AI agents and developers before they apply breaking changes.

graph TD
    subgraph Evidence Sources
        A1["Codebase & Git History"]
        A2["GitHub PRs & Issues (Webhooks & CLI)"]
        A3["Slack Channel Chat logs (Webhooks & CLI)"]
    end
    
    A1 & A2 & A3 -->|Ingestion & Mining| B["LORE Engine"]
    B -->|Builds| C["5-Layer Knowledge Graph"]
    
    subgraph Knowledge Graph Layers
        C1["L1: Structural AST Symbols (Py, Go, TS)"]
        C2["L2: Semantic Vector Store sqlite-vec"]
        C3["L3: Historical Co-changes & Fragility Scores"]
        C4["L4: Decisional Links to ADRs & PRs"]
        C5["L5: Institutional Policy & Boundary Rules"]
    end
    
    C --> C1 & C2 & C3 & C4 & C5
    C -->|Exposes Context| D["Model Context Protocol Server"]
    C -->|Validates Diff| E["LORE Guardian & SARIF Output"]
    
    D -->|Guide Agent| F["Cursor / Claude Desktop / Claude Code"]
    E -->|Block Breaking PR| G["Pull Request Gatekeeper"]
Loading

⚡ Quick Start: Experience LORE in 60 Seconds

1. Install LORE

pip install lore-kg

2. Initialize Workspace & Index Codebase

Set your LLM API key (e.g. Anthropic, OpenAI, DeepSeek, or OpenRouter):

export ANTHROPIC_API_KEY="your-api-key"

Then, run the bootstrap helper inside your repository to scan files and build your Knowledge Graph:

lore init .

3. Run Architectural Audit in CI/CD or PRs

Audit local modifications or PR commit ranges:

lore gh-check --commit-range "origin/main...HEAD" --format sarif --fail-on critical

4. Query the Knowledge Graph

Ask questions about why the codebase is structured the way it is:

lore query "Why did we replace JWT with opaque tokens in auth.py?"

⚖️ What Makes LORE Different?

Feature Standard RAG / Search Modern AI IDEs (Cursor, Cody, Copilot) LORE
AST Symbol Indexing ❌ (Text chunking) ⚠️ Single-file AST / SCIP chunks ✅ 5-Layer Graph + LoD Skeletonization
Architectural Intent & ADRs (L4) ✅ Scoped Decisional Links (L4)
Mined Symbol Co-Change Rules ✅ 800+ Mined Association Rules (L3)
Boundary Condition Miner ✅ Operator Weakening Alerts (> $\rightarrow$ >=)
Inter-Procedural Taint Graph ✅ Source-to-Sink Dataflow Tracing
Technical Due Diligence Pre-Audit ✅ Key-Person & Bus Factor Mining (lore due-diligence)
DORA Change Risk Evidence ✅ Static Code Evidence Collection (lore dora-report)
EU AI Act Asset Extractor ✅ AST Prompt & HITL Extractor (lore ai-act-report)
Local Offline Vector Search ❌ (Cloud dependent) ⚠️ Supported in select IDEs ✅ Embedded sqlite-vec (C)

🏢 Enterprise Code Evidence & Risk Pre-Audit

LORE assists enterprise risk management, M&A due diligence, and compliance teams by mining static codebase evidence and git history into structured audit reports.

Important

Regulatory Disclaimer: LORE generates static code analysis, git churn, and architectural decision evidence metrics to assist internal engineering and compliance teams. LORE does not provide legal advice or formal legal certification under EU Regulation 2022/2554 (DORA) or EU Regulation 2024/1689 (EU AI Act).

graph LR
    subgraph "LORE Core Engine"
        KG["5-Layer Knowledge Graph + AI Asset Extractor"]
    end

    KG --> DD["lore due-diligence"]
    KG --> DORA["lore dora-report"]
    KG --> AIACT["lore ai-act-report"]

    subgraph "M&A / VC Technical Pre-Audit"
        DD --> R1["Bus Factor & Key-Person Offboarding Risk"]
        DD --> R2["Codebase Health & Maintainability Index"]
        DD --> R3["Hidden Co-Change Coupling Matrix (L3)"]
    end

    subgraph "DORA Change Risk Evidence"
        DORA --> D1["Art. 6: Architectural Intent & ADR Links"]
        DORA --> D2["Art. 9: High-Risk Hotspot Identification"]
        DORA --> D3["Art. 11: Change Impact Analysis & SARIF Logs"]
    end

    subgraph "EU AI Act Static Asset Mining"
        AIACT --> A1["Art. 9: Risk Management System Metrics"]
        AIACT --> A2["Art. 11: Model Cards & System Prompts"]
        AIACT --> A3["Art. 14: AST Human Oversight (HITL) Nodes"]
    end
Loading

📊 1. Technical Due Diligence (lore due-diligence)

Mines git commit history and symbol metrics to assist M&A technical audit teams during engineering reviews:

lore due-diligence --project /path/to/repo --format all --output-dir ./reports
  • Bus Factor & Key-Person Risk: Scans git history to flag files with >70% single-author concentration (offboarding risk).
  • Codebase Health Score (0-100): Evaluates structural maintainability, commit churn, and architectural debt.
  • Hidden Co-Change Coupling Matrix: Uncovers implicit dependencies between decoupled modules discovered from historical co-edits.
  • Deliverables: Dark-mode interactive HTML (due_diligence_report.html), Markdown summary, and structured JSON.

🛡️ 2. DORA Change Risk Evidence Collector (lore dora-report)

Extracts static software architecture and change governance evidence related to DORA (EU Regulation 2022/2554) Articles 6, 9, and 11:

lore dora-report --project /path/to/repo --format all --output-dir ./reports
  • Article 6 (Architectural Intent): Verifies linked Architectural Decision Records (ADRs) across code changes.
  • Article 9 (Protection & Vulnerability Control): Flags high-fragility hotspots and unmitigated taint paths.
  • Article 11 (Change Management): Maps co-change impact analysis via SARIF 2.1.0 PR gatekeepers.
  • Deliverables: Evidence reports (dora_compliance_report.html, .md, .json).

🤖 3. EU AI Act Static Asset Mining (lore ai-act-report)

Scans AST declarations for AI models, prompt templates, and Human-in-the-Loop oversight nodes under EU AI Act (Regulation 2024/1689):

lore ai-act-report --project /path/to/repo --format all --output-dir ./reports
  • Article 9 (Risk Management System): Evaluates AI Risk Category (High-Risk vs. Specific Transparency Risk) based on detected frameworks and domain markers.
  • Article 11 & Annex IV (Model Cards & Technical Evidence): Scans AST for LLM frameworks (OpenAI, Anthropic, LangChain, HuggingFace, vLLM) and system prompts.
  • Article 14 (Human Oversight / HITL): Maps AST function nodes for explicit human confirmation gates (HITL override functions) before critical side-effects.
  • Deliverables: Technical evidence reports (ai_act_compliance_report.html, .md, .json).

🛠️ CLI Command Overview

Command Description
lore init Initialize LORE workspace and index project files (bootstrap).
lore due-diligence Extract Technical Due Diligence & Codebase Health Pre-Audit Metrics for M&A, VC & PE.
lore dora-report Extract DORA (EU 2022/2554) & NIS2 Static Change Risk Evidence Reports.
lore ai-act-report Extract EU AI Act (EU 2024/1689) Static Asset & Human Oversight (HITL) Evidence.
lore gh-check Run PR security & architecture audit with --format [markdown|json|sarif] and --fail-on.
lore reindex Re-compute symbol fragility scores & co-changes across existing Knowledge Graphs.
lore dismiss Suppress a false positive LORE warning for a file or symbol persistent in SQLite.
lore query Query the Knowledge Graph for architectural questions (read-only).
lore adr Generate and index an Architectural Decision Record (ADR) to cure Amnesia.
lore mcp Start the Model Context Protocol (MCP) server for Cursor & Claude Desktop.
lore git-hook Install or uninstall LORE pre-commit git hooks.

🛡️ GitHub Action & SARIF Integration

Integrate LORE Guardian into your GitHub Code Scanning and Security tab via native SARIF 2.1.0 output:

# .github/workflows/lore-audit.yml
name: LORE Security & Architecture Guard

on:
  pull_request:
    branches: [ main ]

jobs:
  lore-guard:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0 # Fetch all history for git mining

      - name: Set up Python
        uses: actions/setup-python@v5
        with:
          python-version: '3.10'

      - name: Install LORE
        run: pip install lore-kg

      - name: Run LORE Audit
        run: lore gh-check --commit-range "origin/main...HEAD" --format sarif --fail-on critical > lore-results.sarif

      - name: Upload SARIF report to GitHub Security Tab
        uses: github/codeql-action/upload-sarif@v3
        with:
          sarif_file: lore-results.sarif

📜 Contributing & License

For development setup instructions, please read CONTRIBUTING.md.

LORE is open-source software licensed under the MIT License.

About

The Local Institutional Memory Layer for AI Coding Agents — A 5-layer Knowledge Graph & Semantic Firewall for Cursor, Claude, and CI/CD.

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages