Mem0 announces the core long-term memory algorithm research: extracting key facts through a two-stage pipeline and updating memory to avoid forgetfulness. Under the LOCOMO benchmark, accuracy is 26% higher than OpenAI's built-in memory, P95 reasoning latency is reduced by 91%, and token consumption is decreased by 90%. The enhanced variant Mem0ᵍ introduces a graph database to capture cross-session entity relationships. From memory retrieval to response, the production end takes only 0.71 seconds, far better than the nearly 10 seconds for full context. The research has been accepted by ECAI, and the code is open-sourced on GitHub.

MeNews

2026-05-21 19:55:18

Abstract generation in progress

ME News Report, April 17 (UTC+8), according to Beating Monitoring, personalized AI memory platform Mem0 recently announced the research results of its core long-term memory algorithm. Experimental data shows that, in the LOCOMO benchmark test, Mem0's response accuracy is 26% higher than OpenAI's built-in memory function, and due to its "fact-based" retrieval mechanism, its P95 reasoning latency has decreased by 91%, and token consumption has been reduced by 90%. The core problem this algorithm addresses is the "forgetfulness" phenomenon of AI agents during long-term interactions. Unlike the brute-force approach of simply expanding the LLM context window, Mem0 adopts a two-stage processing pipeline: in the "extraction stage," the system extracts key facts from the latest conversations, rolling summaries, and historical records; in the "update stage," the system compares against a vector database to perform operations such as addition, updating, deletion conflicts, or ignoring, ensuring the memory bank remains concise and consistent. The research also introduces an enhanced variant, Mem0ᵍ. This version incorporates a graph database structure, transforming extracted facts into labeled nodes and edges to capture complex entity relationships across multiple sessions. In practical production environments, Mem0 can complete the entire process from memory retrieval to response generation in 0.71 seconds, while traditional "full context" methods take nearly 10 seconds. Currently, this research has been accepted by the European Conference on Artificial Intelligence (ECAI), and related code has been open-sourced on GitHub. (Source: BlockBeats)

View Original

This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.

10 Likes

Reward
10
7
8
Share

Comment

Add a comment

NightFlightMint

· 6h ago

In the two-stage extraction of key facts, will information be lost in the first stage?

View OriginalReply0

ApeWithNotes

· 6h ago

From forgetfulness to long-term memory, this pain point is accurately targeted.

View OriginalReply0

BorrowingBuddy

· 6h ago

P95 latency decreased by 91%, and optimizing the long tail is more difficult than improving the average latency.

View OriginalReply0

GateUser-f78f1f3e

· 6h ago

Memory retrieval in 0.71 seconds, can it withstand high concurrency in a production environment?

View OriginalReply0

MintColdBrew

· 6h ago

Using graph databases for cross-session entity relationships is a very clever enhancement.

View OriginalReply0

NightFlightPaperCrane

· 6h ago

OpenAI's built-in memory is indeed useless; finally, someone is doing it properly.

View OriginalReply0

GateUser-4bd1cc87

· 6h ago

What is the specific testing scenario where LOCOMO benchmark is 26% higher than OpenAI?

View OriginalReply0

Trending Topics
View More
#
TradfiTradingChallenge
240.52K Popularity
#
HYPEOutperformsAgain
16.33M Popularity
#
DailyPolymarketHotspot
1.02M Popularity
#
GateSquarePizzaDay
1.71M Popularity
#
SpaceXOfficiallyFilesforIPO
6.38K Popularity

Pinned

Sitemap

Mem0 releases long-term memory architecture research: accuracy surpasses OpenAI by 26%, reasoning latency reduced by 91%

Trending Topics

TradfiTradingChallenge

HYPEOutperformsAgain

DailyPolymarketHotspot

GateSquarePizzaDay

SpaceXOfficiallyFilesforIPO

Pinned