AI agent memory systems see new tools and benchmarking standards

The development of long-term memory systems for AI agents is gaining momentum through both practical tools and standardized evaluation frameworks.

Engram has emerged as a tool designed to provide coding agents with persistent memory. It is a standalone Go binary utilizing an SQLite backend with full-text search capabilities. Engram can be integrated with agents like OpenCode via an MCP server, allowing for regular memory updates and the ability to share documented learnings with development teams through Git or self-hosted services.

To address the difficulty of comparing different AI memory systems, the Agent Memory Leaderboard (AML) has been established. The AML provides a common evaluation environment to separate the performance of a memory system from the influence of specific answer models or prompts. As of August 2026, the competition has seen participation from 136 teams and 67 representative memory frameworks across open-source and commercial text memory tracks.