
Agents that remember: built-in long-term memory for any
Harness Memory is the long-term memory layer built into HarnessRouter, an open-source unified interface for agent harnesses. Every agent gets its own memory automatically—no memory service to wire up and no new integration when the harness changes. Memory is loaded when a session starts, used through tools during each turn, and recorded when the turn ends. Agents on different harnesses (e.g., Claude Code and Codex) can share one memory, and agents serving many users keep each person's records separate.
A memory engine is automatically picked for what the agent does (an agent that only transforms given content, like a translator, gets none), with the choice shown in the console and changeable at any time.
Memory is loaded at session start, used through tools during each turn, and recorded when the turn ends; the agent recalls what mattered before it answers.
Agents on different harnesses can share the same memory—in the demo, a Codex agent saves a plan and a Claude Code agent recalls it in a new session.
Each person gets a memory of their own named on the task, records are kept apart, and a person's memory can be erased on its own. Recalled memory reaches the agent as fenced-off data (with attribution), erase is never offered to agents as a tool, and access is granted per memory.
The Memory page has Data, Schema, and Settings tabs so you can see, edit, or delete what an agent remembers.
Mem0 (short facts, Apache-2.0), Graphiti (a graph of things and relations over time, Apache-2.0), Hindsight (consolidated facts, MIT), and ReMe (Markdown files in folders, Apache-2.0). Any memory service can plug in as a provider via the protocol.
The Memories chapter of the Unified Harness Protocol (UHP): an open definition of what a memory is (a named container with a folder-like tree), who may reach it, and how an agent uses it via tools. Servers advertise it with the memories capability.
Harness Memory can be run on self-hosted infrastructure.
A memory of its own is made with the agent. The engine is picked for what the agent does, and the console shows why; the Change link picks another engine, or none.
Memory is loaded when a session starts, used through tools during each turn, and recorded when the turn ends. The engine keeps what mattered, and the agent recalls it before it answers.
Agents on different harnesses share one memory: switch the memory on for an agent on Codex, the Codex agent saves a plan, and the Claude Code agent recalls it in a new session.
Recalled memory is data, not instructions: it reaches the agent fenced off from its instructions, with who wrote it. Only people erase—Erase is never offered to an agent as a tool. People and agents are members alike, and access is granted per memory.
Agent memory is billed as part of the HarnessRouter plans. Short-term memory is metered per GB-hour (about $0.0613 per GB-month at list rates, beta pricing) and long-term memory is one flat price per memory operation ($0.0019 on every listed plan, with long-term storage included)—both beta pricing subject to change with advance notice. Plans start free with no card needed: Free ($0/month, 30 agent
$0/month
For exploring: top up as needed, 30 agent-work minutes included, 1 concurrent task, 1 workspace, own keys, no markup on model usage.
$20/month
For first integration: 30% off top-up rates, $20 of usage included, unlimited concurrent tasks, 1 workspace, own keys.
$200/month
For live apps: 50% off top-up rates, $200 of usage included, unlimited concurrent tasks, 5 workspaces, own keys.
$1,000/month
For high volume: 70% off top-up rates, $1,000 of usage included, unlimited concurrent tasks, unlimited workspaces, own keys, tracing (beta) included.
Custom
Custom terms and volume pricing; custom usage and workspaces; own keys or ours.
It gives agents long-term memory built into HarnessRouter. Memory is loaded when a session starts, used through tools during each turn, and recorded when the turn ends, so the agent recalls what mattered in any later session.
No. Every agent that needs one gets its own memory automatically, with no memory service to wire up and no new integration when the harness changes. The engine is picked for what the agent does and can be changed—or set to none—at any time.
Yes. Agents on different harnesses can share one memory. In the demo, a Codex agent saves a plan and a Claude Code agent recalls it in a new session.
Short-term memory is metered per GB-hour (about $0.0613 per GB-month at list rates) and long-term memory is $0.0019 per memory operation on every listed plan, with long-term storage included. Both are beta pricing subject to change with advance notice.
Yes—the site lists HarnessRouter as open source with 4.4k GitHub stars, and Harness Memory can also be run on self-hosted infrastructure.
Four: Mem0 (short facts, Apache-2.0), Graphiti (a graph of things and how they relate over time, Apache-2.0), Hindsight (consolidated facts, MIT), and ReMe (Markdown files in folders with links, Apache-2.0). Any memory service can also plug in as a provider through the protocol.
Each person gets a memory of their own named on the task, records are kept apart from everyone else's, and one person's memory can be erased on its own. Recalled memory reaches the agent fenced off from its instructions with attribution, and Erase is never offered to an agent as a tool.
Yes. The Memory page has Data, Schema, and Settings tabs where you can see, edit, or delete what an agent remembers, view the record kinds its engine keeps, and see which agents use the memory.