AI Long-Term Memory (ALTM)

** Agentic Memory System for the Next Generation of AI Agents.**

ALTM is a high-performance, asynchronous memory layer designed to give AI agents a persistent, hierarchical, and self-consolidating long-term memory. Unlike simple vector stores, ALTM implements Agentic Routing (PageIndex), Dynamic Importance Decay, and Automated Memory Merging to ensure your agents remember what matters and forget what doesn't.

---

🏛 Exceptional Architecture

The system is built for extreme reliability and low-latency performance:

• Fully Asynchronous Event Loop: Powered by FastAPI and Motor (Async MongoDB), the architecture ensures zero blocking of the event loop. Every database operation and API call is non-blocking, allowing the system to scale to thousands of concurrent requests.
• Resource Efficiency (OOM-Safe): All batch updates (importance decay, pruning) utilize asynchronous cursor iteration. This prevents memory exhaustion (OOM) even when managing millions of memory nodes for a single user.
• Singleton Pattern Resource Management: Implements thread-safe singleton patterns for HTTPX clients and MongoDB connections, preventing socket exhaustion and ensuring high-performance TCP/TLS reuse.
• Lifespan-Managed Lifecycle: Utilizing FastAPI lifespan handlers for graceful startup/shutdown, ensuring all database connections are cleanly initialized and closed.

🛡 Security & Resilience

ALTM is "battle-hardened" against common production failures:

• Thread-Safe Embedding Cache: A lock-guarded LRU cache prevents race conditions during embedding generation.
• Input Sanitization & Validation: Strict Pydantic models with character limits (OOM protection) and automated MongoDB key sanitization to prevent injection attacks during hierarchical indexing.
• Deterministic Context Retrieval: Advanced sorting with _id tie-breakers ensures conversation history is never jumbled, even during millisecond-parallel message ingestion.
• Atlas-Ready Connectivity: Pre-configured for MongoDB Atlas with dnspython support and optimized Vector Search index definitions.

🚀 Special Features

1. PageIndex (Agentic Routing)

Instead of a flat search, ALTM uses an LLM-driven "Archivist" to categorize memories into a hierarchical map (Categories/Topics). The Agentic Router then narrows down search spaces, significantly improving retrieval accuracy and reducing noise.

2. Automated Memory Consolidation

The system detects "near-duplicate" memories and utilizes an LLM to merge them into a single, cohesive node. This prevents "context stuffing" and keeps the agent's memory lean and high-signal.

3. Dynamic Importance & Decay

Memories aren't static. ALTM implements a Reinforcement/Decay system:

• Reinforce: Similarity to new content boosts a memory's importance.
• Decay: Irrelevant memories fade over time.
• Pruning: Low-importance memories are automatically pruned when a user hits the MAX_DEPTH threshold.

---

🛠 Installation

# Clone the repository
git clone https://github.com/your-repo/ai-long-term-memory.git
cd ai-long-term-memory

# Install production dependencies
pip install -r requirements.txt

# Configure your environment
cp sample.env .env
# Add your MONGODB_URI and OPENROUTER_API_KEY

🤝 Call for Open-Source Contributors

We are building the standard for agentic memory, and we want your help. ALTM is currently seeking contributors to help expand in the following areas:

• Multi-Modal Support: Extending memory nodes to store and retrieve image/audio embeddings.
• Graph-Based Retrieval: Integrating Knowledge Graph links between memory nodes.
• Local LLM Providers: Adding first-class support for Ollama and vLLM for fully local deployments.
• Client SDKs: Building TypeScript/Python SDKs for easy integration.

Ready to contribute?

1. Check the Issues tab for good-first-issue tags.
2. Fork the repo and create a feature branch.
3. Submit a PR. We value clean code, async-first patterns, and comprehensive tests.

---

License: MIT
Author: Fahad Siddique Faisal
Status: Production-Ready