🧠 LOB Brain

Lambda-Obsidian Brain — AI Memory Engine with λ-Decay Algorithm

What if your AI Agent never forgot anything important — and automatically forgot what's no longer relevant?

---

🔥 The Problem

Every AI Agent has amnesia. Each session starts from scratch. You re-explain the same context, lose architectural decisions, and watch your token bill climb with redundant prompting.

💡 The Solution

LOB Brain is a local-first memory engine that gives AI Agents a persistent, self-managing memory — backed by SQLite and synced to your Obsidian vault as human-readable Markdown.

AI Agent  ←──  LOB Brain Engine  ←──→  SQLite DB (brain.db)
                                  ←──→  Obsidian Vault (.md files)
                                  ←──→  Gemini / Local TF-IDF Embeddings

---

🧮 The Algorithm — λ-Decay Scoring

LOB Brain's core is a λ-decay memory scoring system, inspired by physics of radioactive decay and Ebbinghaus's forgetting curve:

$$\text{Score} = \text{Importance} \times e^{-\lambda \cdot \Delta t}$$

Where:

• Importance — rated 1–5 when storing (you decide what matters)
• λ (lambda) — decay constant controlling how fast memories fade
• Δt — hours elapsed since last access (reading resets the clock)

Each memory is assigned a fidelity tier based on its live score:

Score	Tier	Description
> 3.5	🔥 HOT	Full text — high relevance, recently used
2.0–3.5	🌡️ WARM	Summary only
1.0–2.0	🧊 COOL	Essence (≤15 words)
0.1–1.0	🌫️ FADED	Hash reference only
< 0.1	💨 GONE	Excluded from context entirely

When assembling context for an AI prompt, memories are loaded at their appropriate fidelity tier, fitting the maximum amount of useful information into a fixed token budget — automatically.

💡 Credit: The λ-decay approach for AI memory management was inspired by the Temm1e project by temm1e-labs. We are grateful for their foundational research on adaptive memory decay models for AI systems.

---

📦 What's in This Repo (Core Engine)

This repository contains the open-source brain core library:

Module	Description
src/database.ts	SQLite operations — store, query, recall, GC, FTS5 full-text search
src/decay.ts	λ-decay scoring, fidelity tier selection, token-budget context assembly
src/embeddings.ts	Gemini text-embedding-004 + local TF-IDF fallback
src/embed_stats.ts	Gemini API usage tracking (calls, estimated tokens)
src/clustering.ts	K-means clustering of memories by topic similarity
src/obsidian.ts	Bidirectional sync with Obsidian vault (Markdown ↔ SQLite)
src/ingestion.ts	Chunked ingestion of Markdown/TXT files for RAG
src/types.ts	Shared TypeScript type definitions
tests/	Unit tests for core logic (27 passing)

---

🔌 MCP Integration Layer

The MCP server that exposes these capabilities as AI tools (for Claude Desktop, Antigravity, Cursor, etc.) is available separately.

It includes:

• 10 ready-to-use MCP tools (brain_store, brain_query, brain_recall, brain_context, brain_clusters...)
• SSE server for multi-machine / team setups
• Web dashboard for memory management
• Project-scoped API key access control

Want the full MCP integration? Contact me:

---

🚀 Getting Started (Core Library)

git clone https://github.com/ankmoon/lob-brain.git
cd lob-brain
npm install
npm run build

Using the Core in Your Project

import { BrainDatabase } from './src/database.js';
import { assembleContext, decayScore } from './src/decay.js';
import { getEmbedding } from './src/embeddings.js';

// Initialize the brain
const db = new BrainDatabase('./brain.db');

// Store a memory
db.store({
  hash: 'abc123',
  full_text: 'Decided to use PostgreSQL for the auth service',
  summary: 'PostgreSQL chosen for auth service',
  essence: 'PostgreSQL for auth — decided 2026-03',
  importance: 4,
  project: 'my-project',
  memory_type: 'decision',
  sensitivity: 'public',
  tags: ['database', 'auth', 'architecture'],
  owner_id: 'default',
  created_at: new Date().toISOString(),
  last_accessed: new Date().toISOString(),
  access_count: 0,
});

// Query memories
const results = db.ftsSearch('PostgreSQL', 'my-project', 10);

// Get context block (respects token budget)
const candidates = db.queryCandidates('my-project', 50);
const context = assembleContext(candidates, 8000);
console.log(context); // Ready to inject into system prompt

Run Tests

npm test
# 27 tests, 0 failures

---

⚙️ Environment Variables

Variable	Description
GEMINI_API_KEY	Optional — enables Gemini text-embedding-004 for semantic search. Falls back to local TF-IDF if not set. Free tier: 1500 req/day at aistudio.google.com
LAMBDA_BRAIN_DB_PATH	Optional — custom path to SQLite database (default: ./brain.db)

---

🤝 Contributing

Contributions to the core engine are welcome!

1. Fork the repo
2. Create a feature branch: git checkout -b feature/my-feature
3. Run tests: npm test
4. Open a Pull Request

---

📄 License

MIT — Core library is free to use, modify, and build upon.

---

Built with ❤️ for the AI Agent community.
Special thanks to temm1e-labs/temm1e for the λ-decay memory model that powers this project.

---

ghostcreator