memrust: A Unified Memory Engine for AI Agents

Introduction

memrust is a specialized tool designed to give AI agents a true, shared memory. While most AI systems rely on simple vector databases that struggle with precise details, memrust acts as a unified brain. It combines different types of data signals to help agents remember facts, relationships, and recent events accurately. Instead of just finding similar text, it understands exact terms, connections between ideas, and how fresh the information is.

Benefits

memrust offers several key advantages over traditional memory systems. First, it provides much higher accuracy. While standard vector search might only find the right answer 27% of the time for specific codes or identifiers, memrust achieves 75% recall by using a hybrid approach. This system fuses four distinct signals: semantic meaning, exact text matching, relationship graphs, and time-based recency. Second, it simplifies infrastructure. Teams no longer need to build a complex stack of separate tools for indexing, scheduling, and access control. memrust handles all of this in one binary. Third, it ensures reliability. The system is built with durability in mind, allowing it to recover data even after unexpected crashes. Finally, it supports multiple agents sharing knowledge through a single endpoint, enabling better collaboration between different AI tools.

Use Cases

memrust is ideal for teams building complex AI agent workflows. It works well with frameworks like LangGraph, CrewAI, and various coding assistants such as Claude Code and Cursor. Developers can use it to store episodic memories, like what happened during a software deployment, or semantic knowledge about specific error codes. The tool supports both self-hosted deployment for complete control and a hosted private beta for teams that prefer a managed service. It is particularly useful when an application needs to distinguish between very similar strings, such as different error numbers, or when it needs to prioritize recent information over older data. The system also provides detailed scores for each type of match, helping developers understand why the agent retrieved specific information.

Pricing

memrust offers flexible options for different needs. The self-hosted version is free and has no usage limits. Users can run it as a lightweight Docker image or install the binary directly on their servers. For teams that prefer a managed solution, there is a hosted private beta available. This option includes features like project management, usage metering, and audit logs. The free tier for the hosted service includes up to three projects and 25,000 requests per month.

Vibes

As a new tool in the AI space, memrust is currently in a private beta phase for its hosted service. The project is open-source and licensed under Apache-2.0, which has generated interest among developers who want to self-host the solution. Early benchmarks show significant improvements in recall rates compared to existing vector databases, suggesting strong potential for adoption in production environments. The focus on multi-agent visibility and durability indicates a design that aligns well with the growing demand for reliable AI infrastructure.

Additional Information

memrust is written in the Rust programming language. This choice ensures high performance, predictable latency, and efficient memory usage without the pauses often caused by garbage collection. The project is available as a static binary, a Docker image, and through SDKs for Python and TypeScript. It integrates with the Model Context Protocol, allowing it to work seamlessly with various agent frameworks. The system supports multi-tenant architectures, meaning different users or agents can have isolated memory spaces within the same engine.