🧠 AgentRank

Temporal-aware embeddings for AI agent memory retrieval.

---

The Problem

Standard embedding models (OpenAI, Cohere, MiniLM) treat "yesterday" and "6 months ago" identically. For AI agents with long-term memory, this breaks temporal reasoning completely.

AgentRank solves this with embeddings that understand:

• ⏰ When memories happened (temporal awareness)
• 📂 What type of memory it is (episodic, semantic, procedural)
• ⚡ 21% better retrieval on agent memory benchmarks

---

Installation

pip install agentrank

---

Quick Start

from agentrank import AgentRankEmbedder

# Load model
model = AgentRankEmbedder.from_pretrained("vrushket/agentrank-base")

# Encode with temporal context
embeddings = model.encode(
    texts=["User prefers Python for backend development"],
    temporal_info=[7],        # 7 days ago
    memory_types=["semantic"] # It's a preference
)

# Use embeddings for retrieval
print(embeddings.shape)  # [1, 768]

---

Model Family

Model	Type	Params	Use Case	HuggingFace
AgentRank-Base	Embedder	149M	Best quality retrieval	vrushket/agentrank-base
AgentRank-Small	Embedder	33M	Fast inference	vrushket/agentrank-small
AgentRank-Reranker	Cross-encoder	149M	Accurate reranking	vrushket/agentrank-reranker

---

Two-Stage Retrieval Pipeline

For best results, use a two-stage pipeline:

from agentrank import AgentRankEmbedder
from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

# Stage 1: Fast retrieval with embedder
embedder = AgentRankEmbedder.from_pretrained("vrushket/agentrank-base")
query_embedding = embedder.encode(["What's my Python preference?"])
# ... search vector DB → get top-50 candidates ...

# Stage 2: Accurate reranking with cross-encoder
reranker = AutoModelForSequenceClassification.from_pretrained("vrushket/agentrank-reranker")
tokenizer = AutoTokenizer.from_pretrained("vrushket/agentrank-reranker")

def rerank(query, candidates, top_k=10):
    scored = []
    for memory in candidates:
        inputs = tokenizer(query, memory, return_tensors="pt", truncation=True)
        with torch.no_grad():
            score = torch.sigmoid(reranker(**inputs).logits).item()
        scored.append((score, memory))
    return sorted(scored, reverse=True)[:top_k]

top_10 = rerank("What's my Python preference?", top_50_candidates)

---

Benchmarks

Model	MRR	Recall@1	Recall@5	NDCG@10
AgentRank-Base	0.6496	0.4440	99.6%	0.6786
AgentRank-Small	0.6375	0.4460	97.4%	0.6797
MPNet-base-v2	0.5351	0.3660	79.6%	0.6335
MiniLM-L6-v2	0.5297	0.3720	75.2%	0.6370

+22% MRR improvement over baseline embedding models.

Reranker	Validation Accuracy	Val Loss
AgentRank-Reranker	89.11%	0.2554

---

Key Features

Temporal Embeddings

10 learnable time buckets encode recency:

• Today/Yesterday, This Week, This Month, Last Quarter, etc.
• Model learns what "recent" means in context

Memory Type Embeddings

Distinguish between:

• Episodic: Events ("We discussed Python yesterday")
• Semantic: Facts/preferences ("User likes Python")
• Procedural: Instructions ("To deploy, run npm build")

Architecture

• Base: ModernBERT-base / MiniLM
• Temporal + Type embeddings added (scaled by 0.1)
• Trained on 500K synthetic agent memory samples
• Hard negative mining with 7 negative types

---

Works Great With

CogniHive — Multi-agent memory with "who knows what" routing

pip install cognihive

Together: CogniHive routes questions to the right agent, AgentRank retrieves the right memories.

---

Links

• GitHub: github.com/vmore2/AgentRank-base
• HuggingFace: huggingface.co/vrushket
  • agentrank-base
  • agentrank-small
  • agentrank-reranker
• CogniHive: pypi.org/project/cognihive

---

Contact

• Author: Vrushket More
• Email: vrushket2604@gmail.com
• LinkedIn: linkedin.com/in/vrushket-more-07b38b25b
• Issues: GitHub Issues

---

License

Apache 2.0 — Free for commercial use.