agentrank 0.1.0

Temporal-aware embedding model for AI agent memory retrieval

🧠 AgentRank

The First Retrieval Model Family Optimized for AI Agent Memory

Every existing agent memory system (Mem0, Letta, CogniHive, etc.) uses generic embeddings. AgentRank is the first specialized solution.

Model Family

Model	Size	Base	Use Case
agentrank-small	33M	MiniLM	Edge/fast
agentrank-base	110M	ModernBERT-base	Standard
agentrank-reranker	140M	Cross-encoder	Top-k reranking

Novel Features

1. Temporal Position Embeddings - Understands when memories were created
2. Memory Type Embeddings - Distinguishes episodic/semantic/procedural
3. Importance Prediction - Auxiliary task for ranking critical memories

Quick Start

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("vrushank/agentrank-base")

# Encode memories
memories = [
    "User prefers dark mode in all applications",
    "Yesterday we discussed Python debugging",
    "To deploy: run pytest → build docker → push to ECR"
]
embeddings = model.encode(memories)

# Encode query and find relevant memories
query = "What are the user's UI preferences?"
query_emb = model.encode(query)
similarities = model.similarity(query_emb, embeddings)

Project Structure

agentrank/
├── data/
│   ├── generators/      # Memory & query generators
│   └── datasets/        # Generated training data
├── models/
│   ├── embedder.py      # AgentRank embedder
│   └── reranker.py      # AgentRank reranker
├── training/
│   ├── train_embedder.py
│   └── train_reranker.py
├── evaluation/
│   └── agentmembench.py
└── scripts/
    ├── generate_data.py
    └── upload_to_hub.py

Training Data

• 500K synthetic agent memory traces
• 3 memory types: Episodic (40%), Semantic (35%), Procedural (25%)
• 5 types of hard negatives for robust training

License

Apache 2.0