Harness Memory

Automated and Cross-Platform Memory Management

Harness Memory automatically provides each agent that requires memory with its own dedicated long-term storage, eliminating the need for additional memory service integration or reconfiguration when switching platforms (referred to as harnesses). The memory is loaded at the start of each session, available to the agent during interaction turns, and updated after each turn ends. This system ensures agents recall relevant information before responding to new tasks.

Moreover, agents operating on different harnesses can share a single memory space. For example, in a demonstration, a Claude Code agent and a Codex agent simultaneously access the same memory, enabling cross-agent information sharing and collaboration.

Multiple Memory Engines for Diverse Needs

The system comes with four memory engines, each managing and storing data differently: Mem0 stores short facts extracted from conversations; Graphiti builds a relational graph over time; Hindsight consolidates important facts in the background; ReMe manages linked Markdown files in folders. More engines are planned to provide flexible and diverse memory management options.

Privacy and Multi-User Memory Separation

To address privacy concerns when an agent serves multiple users, Harness Memory creates a separate memory for each user, tagged by task. This design keeps user records isolated and allows for individual memory deletion, maintaining data privacy and security.