Current State of Memory Technology in Agent Harnesses

The Agent Harness, the environment in which AI software actually runs, is evolving beyond a simple model into a core product that handles context management, tool orchestration, and memory management.

Memory is broadly divided into three layers.

• Working memory: Information that exists within the context window during a session, reset when the session ends.
• External memory: Persistent storage outside the session, such as vector stores and knowledge graphs, currently used in most production environments.
• Parametric memory: Knowledge encoded in the model's weights.

The implementation methods of major agent harnesses are as follows.

Claude Code (Anthropic) utilizes user-written CLAUDE.md and agent-written Auto-memory. However, since it selects information based on file names rather than semantic search, there is a limitation in that even highly relevant files may not be selected if their names are not appropriate.

Managed Agents (Anthropic) is a hosted runtime that manages memory through /mnt/memory/, a form of immutable version file system. This is optimized for multi-agent collaboration and audit, but has limitations for long-term personal memory management.

Codex (OpenAI) is based on markdown files, with a write path consisting of a first stage that extracts information after a session ends and a second stage that consolidates it.