The man behind Cursor's "memory" feature

2025-07-24 · Source: Greg Kamradt · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Advanced, long

Summary

Yash from Cursor's engineering team detailed the development of the IDE's "Memory" feature, a load-bearing process for the $9 billion IDE. This feature aims to augment three types of agent context: directional (narrowing search), operational (runbook conventions), and behavioral (user preferences). The team prototyped two main approaches: a "sidecar" model that independently processes conversations to decide what to save or update, and a "tool call" approach where the main model uses an "update memory" tool. While Sonnet 4 and Opus 4 models showed improved instruction following for tool calls, they tended to generate task-specific memories, which are less desirable for a coding agent. The sidecar approach, despite initial challenges with filtering task-specific information, proved effective with newer reasoning models, requiring less aggressive prompting. Evaluation focused on memory quality, particularly filtering out task-specific data, rather than retrieval speed. Cursor currently keeps memory generation somewhat hidden from users to manage expectations about memory perfection.

Key takeaway

For AI Engineers developing coding assistants, understanding the distinction between task-specific and generalizable memories is crucial. You should prioritize filtering out transient conversational data to prevent the model from "doubling down" on incorrect or irrelevant information, which can significantly degrade user experience. Consider a hidden memory generation process to manage user expectations, as models can effectively filter noise from imperfect memory banks.

Key insights

Cursor's Memory feature enhances agent context by distinguishing between task-specific and generalizable information for improved coding assistance.

Principles

• Context breakdown clarifies agent behavior.
• User control is paramount in coding agents.
• Memory quality trumps quantity.

Method

Cursor prototyped "sidecar" (background model decision) and "tool call" (main model with update tool) approaches for memory generation, prioritizing filtering task-specific data and leveraging advanced reasoning models.

In practice

• Filter task-specific memories in coding agents.
• Conceal memory generation from users.
• Use advanced models for precise prompting.

Topics

• AI Agent Memory
• Context Management
• Coding Agents
• Sidecar Architecture
• Large Language Models

Best for: Machine Learning Engineer, AI Engineer, Software Engineer

Related on AIssential

• Engineering Memory for AI Agents: A Practical Beginner’s Guide — AI agent memory selectively retains past interaction details to enable contextual, multi-step task execution.
• How Chatbots Remember Who You Are — Chatbot memory, both short-term and long-term, is essential for maintaining conversational context and enhancing user experience.
• AI Agents Have Amnesia. A Bigger Context Window Won’t Cure It. — AI agents require a dedicated, architected memory layer to overcome inherent amnesia, as context windows are buffers, not durable truth models.
• GPT-Realtime-2, Directionally Bad and Agent Memory — Agent memory and context engineering are crucial for enhancing AI agent performance and reducing operational costs.
• OCR-Memory: Optical Context Retrieval for Long-Horizon Agent Memory — Visual encoding of agent trajectories enables long-horizon memory with reduced token costs and hallucination.
• How AI Agents Manage Memory and Avoid Forgetfulness — The perceived memory of AI agents is an engineered system around stateless LLMs, not an inherent model capability.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Greg Kamradt.