Reranker-Enhanced Search - Mem0

Reranker-enhanced Search

Reranker-enhanced search adds a second scoring pass after vector retrieval so Mem0 can return the most relevant memories first. Enable it when keyword similarity alone misses nuance or when you need the highest-confidence context for an agent decision.

You’ll use this when…

• Queries are nuanced and require semantic understanding beyond vector distance.
• Large memory collections produce too many near matches to review manually.
• You want consistent scoring across providers by delegating ranking to a dedicated model.

Reranking raises latency and, for hosted models, API spend. Benchmark with production traffic and define a fallback path for latency-sensitive requests.

The Configure it and See it in action snippets below use the Python SDK. The self-hosted TypeScript SDK supports the Cohere, Zero Entropy, Sentence Transformer, Hugging Face, and LLM rerankers; see TypeScript SDK.

TypeScript SDK

The self-hosted TypeScript SDK (mem0ai/oss) ships five rerankers: Cohere, Zero Entropy, Sentence Transformer, Hugging Face, and the LLM reranker. Configure one under reranker, then opt in per search with rerank: true. Keys are camelCase (apiKey, not api_key).Provider SDKs are peer dependencies. Install the one your reranker needs:

pnpm add cohere-ai          # cohere
pnpm add zeroentropy        # zero_entropy
pnpm add @huggingface/transformers   # sentence_transformer, huggingface
# llm_reranker defaults to openai (already a core dependency); install another
# provider's SDK only if you nest a different one under config.llm

Hosted Rerankers (Cohere, Zero Entropy)

Both call a hosted API and read their key from config or the provider’s environment variable (COHERE_API_KEY, ZERO_ENTROPY_API_KEY).

import { Memory } from "mem0ai/oss";

// Cohere reranker (defaults to the rerank-v3.5 model)
const memory = new Memory({
  reranker: {
    provider: "cohere",
    config: { apiKey: process.env.COHERE_API_KEY },
  },
});

const results = await memory.search("What are my food preferences?", {
  filters: { userId: "alice" },
  rerank: true,
});

// Zero Entropy reranker (defaults to the zerank-1 model)
const memory = new Memory({
  reranker: {
    provider: "zero_entropy",
    config: { apiKey: process.env.ZERO_ENTROPY_API_KEY },
  },
});

Local Cross-Encoders (Sentence Transformer, Hugging Face)

Both run a cross-encoder locally with Transformers.js: no API key, no network at inference time. Because Transformers.js runs ONNX weights, the default models are the ONNX mirrors of the Python SDK’s defaults (sentence_transformer → Xenova/ms-marco-MiniLM-L-6-v2, huggingface → Xenova/bge-reranker-base). Point model at any ONNX-exported cross-encoder on the Hub to override.

const memory = new Memory({
  reranker: {
    provider: "sentence_transformer", // or "huggingface"
    config: {
      // model: "Xenova/bge-reranker-base",  // override the default
      device: "cpu", // Transformers.js device: "cpu" | "wasm" | "webgpu"
      maxLength: 512, // max tokens per query-document pair
      normalize: true, // sigmoid-normalize logits to [0, 1] (default)
    },
  },
});

const results = await memory.search("What movies do I like?", {
  filters: { userId: "alice" },
  rerank: true,
});

batchSize and showProgressBar are accepted for config parity with the Python SDK but are no-ops in this runtime, because a memory search reranks a small candidate set in a single in-process forward pass. The model is downloaded once and cached in-process on first use.

LLM Reranker

To score with an LLM instead of a dedicated reranker, use the llm_reranker provider. It builds its own LLM from the reranker’s config (defaulting to openai / gpt-4o-mini) rather than reusing the Memory’s main llm:

const memory = new Memory({
  reranker: {
    provider: "llm_reranker",
    config: { apiKey: process.env.OPENAI_API_KEY },
  },
});

const results = await memory.search("What movies do I like?", {
  filters: { userId: "alice" },
  rerank: true,
});

Nest a different provider under config.llm to override the default:

const memory = new Memory({
  reranker: {
    provider: "llm_reranker",
    config: {
      llm: {
        provider: "anthropic",
        config: { apiKey: process.env.ANTHROPIC_API_KEY },
      },
    },
  },
});

Config Reference

rerank is opt-in per search and a no-op when no reranker is configured. If the reranker call fails, Mem0 logs a warning and returns the original vector-ranked results.

Feature Anatomy

• Initial vector search: Retrieve candidate memories by similarity.
• Reranker pass: A specialized model scores each candidate against the original query.
• Reordered results: Mem0 sorts responses using the reranker’s scores before returning them.
• Optional fallbacks: Toggle reranking per request or disable it entirely if performance or cost becomes a concern.

Supported Providers

• Cohere: Multilingual hosted reranker with API-based scoring.
• Sentence Transformer: Local Hugging Face cross-encoders for GPU or CPU.
• Hugging Face: Bring any hosted or on-prem reranker model ID.
• LLM Reranker: Use your preferred LLM (OpenAI, etc.) for prompt-driven scoring.
• Zero Entropy: High-quality neural reranking tuned for retrieval tasks.

Provider Comparison

Basic Setup

from mem0 import Memory

config = {
    "reranker": {
        "provider": "cohere",
        "config": {
            "model": "rerank-v3.5",
            "api_key": "your-cohere-api-key"
        }
    }
}

m = Memory.from_config(config)

Confirm results["results"][0]["score"] reflects the reranker output: if the field is missing, the reranker was not applied. Set top_k to the smallest candidate pool that still captures relevant hits. Smaller pools keep reranking costs down.

Provider-Specific Options

# Cohere reranker
config = {
    "reranker": {
        "provider": "cohere",
        "config": {
            "model": "rerank-v3.5",
            "api_key": "your-cohere-api-key",
            "top_k": 10,
            "return_documents": True
        }
    }
}

# Sentence Transformer reranker
config = {
    "reranker": {
        "provider": "sentence_transformer",
        "config": {
            "model": "cross-encoder/ms-marco-MiniLM-L-6-v2",
            "device": "cuda",
            "max_length": 512
        }
    }
}

# Hugging Face reranker
config = {
    "reranker": {
        "provider": "huggingface",
        "config": {
            "model": "BAAI/bge-reranker-base",
            "device": "cuda",
            "batch_size": 32
        }
    }
}

# LLM-based reranker
config = {
    "reranker": {
        "provider": "llm_reranker",
        "config": {
            "provider": "openai",
            "model": "gpt-5-mini",
            "api_key": "your-openai-api-key",
            "top_k": 5
        }
    }
}

Keep authentication keys in environment variables when you plug these configs into production projects.

Full Stack Example

config = {
    "vector_store": {
        "provider": "qdrant",
        "config": {
            "host": "localhost",
            "port": 6333
        }
    },
    "llm": {
        "provider": "openai",
        "config": {
            "model": "gpt-5-mini",
            "api_key": "your-openai-api-key"
        }
    },
    "embedder": {
        "provider": "openai",
        "config": {
            "model": "text-embedding-3-small",
            "api_key": "your-openai-api-key"
        }
    },
    "reranker": {
        "provider": "cohere",
        "config": {
            "model": "rerank-v3.5",
            "api_key": "your-cohere-api-key",
            "top_k": 15,
            "return_documents": True
        }
    }
}

m = Memory.from_config(config)

A quick search should now return results with both vector and reranker scores, letting you compare improvements immediately.

Async Support

from mem0 import AsyncMemory

async_memory = AsyncMemory.from_config(config)

async def search_with_rerank():
    return await async_memory.search(
        "What are my preferences?",
        filters={"user_id": "alice"},
        rerank=True
    )

import asyncio
results = asyncio.run(search_with_rerank())

Inspect the async response to confirm reranking still applies; the scores should match the synchronous implementation.

Tune Performance and Cost

# GPU-friendly local reranker configuration
config = {
    "reranker": {
        "provider": "sentence_transformer",
        "config": {
            "model": "cross-encoder/ms-marco-MiniLM-L-6-v2",
            "device": "cuda",
            "batch_size": 32,
            "top_k": 10,
            "max_length": 256
        }
    }
}

# Smart toggle for hosted rerankers
def smart_search(query, user_id, use_rerank=None):
    if use_rerank is None:
        use_rerank = len(query.split()) > 3
    return m.search(query, filters={"user_id": user_id}, rerank=use_rerank)

Use heuristics (query length, user tier) to decide when to rerank so high-signal queries benefit without taxing every request.

Handle Failures Gracefully

try:
    results = m.search("test query", filters={"user_id": "alice"}, rerank=True)
except Exception as exc:
    print(f"Reranking failed: {exc}")
    results = m.search("test query", filters={"user_id": "alice"}, rerank=False)

Always fall back to vector-only search: dropped queries introduce bigger accuracy issues than slightly less relevant ordering.

Migrate from v0.x

# Before: basic vector search
results = m.search("query", filters={"user_id": "alice"})

# After: same API with reranking enabled via config
config = {
    "reranker": {
        "provider": "sentence_transformer",
        "config": {
            "model": "cross-encoder/ms-marco-MiniLM-L-6-v2"
        }
    }
}

m = Memory.from_config(config)
results = m.search("query", filters={"user_id": "alice"})

See It in Action

Basic Reranked Search

results = m.search(
    "What are my food preferences?",
    filters={"user_id": "alice"}
)

for result in results["results"]:
    print(f"Memory: {result['memory']}")
    print(f"Score: {result['score']}")

Expect each result to list the reranker-adjusted score so you can compare ordering against baseline vector results.

Toggle Reranking Per Request

results_with_rerank = m.search(
    "What movies do I like?",
    filters={"user_id": "alice"},
    rerank=True
)

results_without_rerank = m.search(
    "What movies do I like?",
    filters={"user_id": "alice"},
    rerank=False
)

Log the reranked vs. non-reranked lists during rollout so stakeholders can see the improvement before enforcing it everywhere.

Combine With Metadata Filters

results = m.search(
    "important work tasks",
    filters={
        "AND": [\
            {"user_id": "alice"},\
            {"category": "work"},\
            {"priority": {"gte": 7}}\
        ]
    },
    rerank=True,
    top_k=20
)

Verify filtered reranked searches still respect every metadata clause: reranking only reorders candidates; it never bypasses filters.

Real-World Playbooks

Customer Support

config = {
    "reranker": {
        "provider": "cohere",
        "config": {
            "model": "rerank-v3.5",
            "api_key": "your-cohere-api-key"
        }
    }
}

m = Memory.from_config(config)

results = m.search(
    "customer having login issues with mobile app",
    filters={"agent_id": "support_bot", "category": "technical_support"},
    rerank=True
)

Top results should highlight tickets matching the login issue context so agents can respond faster.

Content Recommendation

results = m.search(
    "science fiction books with space exploration themes",
    filters={"user_id": "reader123", "content_type": "book_recommendation"},
    rerank=True,
    top_k=10
)

for result in results["results"]:
    print(f"Recommendation: {result['memory']}")
    print(f"Relevance: {result['score']:.3f}")

Expect high-scoring recommendations that match both the requested theme and any metadata limits you applied.

Verify The Feature Is Working

• Inspect result payloads for both score (vector) and reranker scores; mismatched fields indicate the reranker didn’t execute.
• Track latency before and after enabling reranking to ensure SLAs hold.
• Review provider logs or dashboards for throttling or quota warnings.
• Run A/B comparisons (rerank on/off) to validate improved relevance before defaulting to reranked responses.