DeepSeek alum launches open-source personal AI agent for local PCs
Linger allows users to run high-parameter models on consumer hardware while keeping all personal data on local devices.
The short version
- Linger operates directly on PC web pages, emails, and calendars without needing extra login configurations.
- The system achieves 95% memory accuracy in long-term conversational tests, outperforming several existing memory-augmented AI systems.
- A 125B parameter model can run on an RTX 3060 graphics card with 12GB VRAM at 45.5 tokens per second.
- The project is fully open-source and emphasizes local ownership of data and intelligence over cloud-based token leasing.
Why it matters
This marks a shift toward 'distributed intelligence' where users can own their AI's memory and data locally. It enables high-end AI capabilities on affordable hardware, removing the need for recurring subscription fees and cloud privacy risks.
A new home for local AI
Tony Zhang, a former project lead at DeepSeek who worked on the DeepSeek-VL vision-language model, has launched a new startup project called Linger. The project focuses on creating a personal AI agent that lives directly on a user's personal computer rather than in the cloud. This approach is designed to give users full ownership of their personal context, data, and the intelligence that powers their daily tasks.
Unlike many competitors that prioritize building mobile applications, Linger focuses on the PC platform. This allows the agent to interact directly with existing web pages, email inboxes, and calendars already logged into the user's computer. By operating within these existing environments, the system avoids the need for users to configure a large number of separate login credentials for different services.
The project is being released as an open-source system, following a philosophy that intelligence should be distributed. Zhang argues that while current AI models are trained in data centers and leased as tokens, personal agents should be decentralized. This ensures that a user's private information never has to leave their local device to be processed by a third party.
Memory and proactive habits
At the heart of Linger is a system called 'Another Me,' which functions as a comprehensive memory hub for the user. The agent locally reads through various sources such as chat records, meeting notes, and calendars to organize information into a personal knowledge base. This knowledge base is updated every night so that all other agents within the system can query the user's history and preferences.
In testing, Linger reached a 95% accuracy rate in the LoCoMo benchmark, which measures long-term conversational memory. This performance is reportedly higher than several other memory-augmented AI systems currently on the market. Because it can recall specific details from the past, the agent can act more proactively by observing the user's information flow and deciding when to provide reminders based on urgency.
The system also features 'implicit reinforcement' to learn from human feedback over time. For example, if a user tells the agent they received an invoice last week and wanted a reminder immediately, the system records that preference. If the user ignores a specific type of notification twice, the agent learns to stay silent on those topics in the future. This allows the assistant to become increasingly familiar with specific work habits without requiring manual rule programming for every single interaction.
Running large models on small hardware
One of the most significant technical hurdles for local AI is the amount of video memory (VRAM) required to run large models. Linger addresses this using a specialized MoE (Mixture of Experts) inference engine called PolyStrata. This engine acts as an intelligent tiered storage mechanism that moves data dynamically between the GPU, system memory, and the solid-state drive (SSD).
By placing the most active 'expert modules' in the fastest GPU memory and moving less frequent ones to the SSD, the system can run models that would otherwise be too large for consumer hardware. As a result, a computer with an RTX 3060 graphics card and 12GB of VRAM can run a 125B parameter model at a speed of 45.5 tokens per second. On more powerful hardware like an RTX 5090, that speed can increase to 179 tokens per second.
To simplify the setup for non-technical users, Linger includes a 'Model Autopilot' feature. This tool automatically detects the hardware configuration of the user's device, such as the amount of RAM and GPU power available. It then recommends and installs the best combination of language models, inference engines, and quantization schemes to provide the best possible performance for that specific machine.
The team and the vision
The technical foundation of Linger is supported by a team with deep research backgrounds in computer vision and generative models. Tony Zhang joined the project after serving as an AI scientist at DeepSeek and later as a principal investigator at Zhejiang University. His previous work includes well-known projects like VQ-Diffusion and Rodin 3D digital humans, which highlight the high level of expertise behind the new startup.
The inference engine that powers Linger's local deployment was reportedly built by an 18-year-old prodigy on the team. This highlights the high level of technical innovation being packed into the open-source release. The goal is to provide a truly powerful AI brain that stays on standby 24/7 without any per-token usage fees, effectively making high-end AI accessible to anyone with a standard gaming laptop or MacBook.
While some observers have compared Linger to existing bots from the DeepSeek ecosystem, Zhang clarified the distinction in the official announcement. He noted that while some bots focus on collaboration between different AI agents, Linger is specifically optimized for the interaction between humans and their own personal agents. It is designed to navigate the local environment of the user's computer to perform tasks, rather than just acting as a standalone chat interface.
What to watch
- How the open-source community adopts the PolyStrata inference engine.
- Whether Linger's proactive reminder system can accurately distinguish between high and low-priority tasks without manual tuning.
Try it
Check if your current laptop has at least 12GB of VRAM to see if it can run the 125B parameter model locally.
Sources and links
- Reportingeu.36kr.com: the article we sourced
First reported by eu.36kr.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.