Supported Harnesses
How Agent Memory maps to memory systems in popular agent harnesses.
Agent Memory is designed around patterns already used by stateful agent harnesses. Support varies today: some harnesses implement the format directly, while others already have compatible pieces and can reach full support through configuration or extensions.
Core contract
“Root Markdown” means every flat .md file is loaded into context. “On-demand reads” means the agent can read deferred files when an index or task points to them.
| Harness | Root Markdown | On-demand reads | Freshness |
|---|---|---|---|
| Letta Code | Yes | Yes | Yes |
| Claude Code | Partial | Yes | Yes |
| Codex | Partial | Yes | Yes |
| OpenClaw | Partial | Yes | Yes |
| Hermes Agent | Partial | Partial | Yes |
| Pi | Extension | Extension | Extension |
Partial means the harness implements a narrower version of the behavior, such as loading only specific root Markdown files. Extension means the behavior can be added through the harness's extension system but is not built in.
Optional features
| Harness | Write access | Learned indexes | Root length limit | Frontmatter | Git history |
|---|---|---|---|---|---|
| Letta Code | Yes | Yes | Configurable | Yes | Yes |
| Claude Code | Yes | Yes | 200 lines / 25 KB | Yes | No |
| Codex | Partial | Yes | Token budget | No | Yes |
| OpenClaw | Yes | Partial | Configurable | No | Optional |
| Hermes Agent | Partial | No | Character budget | No | No |
| Pi | Extension | Extension | Extension | Extension | Extension |
These features are optional. A harness can fully implement the core contract with read-only memory and without frontmatter, git, or a fixed length limit.
Letta Code
Full support. Agent Memory is built into Letta Code. Each agent owns a portable, git-backed memory folder. Root Markdown files are loaded into context and directories are available to the agent on demand.
Claude Code
Partial support. Claude Code auto memory uses a MEMORY.md index with topic files read on demand. Its memory is scoped per repository and stored locally, rather than belonging to one agent across projects and machines. A setup guide can bridge the remaining differences with Claude Code configuration and hooks.
Codex
Partial support. Codex maintains a shared ~/.codex/memories/ directory with MEMORY.md, summaries, and detailed rollout files. Its generated memory pipeline already uses a compact index with deeper files available for recall, but it does not yet implement the complete Agent Memory loading contract.
OpenClaw
Partial support. OpenClaw gives each agent a workspace with always-loaded files such as MEMORY.md and USER.md, plus detailed memory files retrieved on demand. Its existing layout maps closely to Agent Memory, though its bootstrap file set and loading rules are OpenClaw-specific.
Hermes Agent
Partial support. Hermes Agent stores durable memory in MEMORY.md and USER.md and loads a frozen snapshot at session start. It does not yet load arbitrary root Markdown files or expose a general progressively disclosed memory tree.
Pi
Extension path. Pi extensions can modify the system prompt before each agent turn and give the model access to a persistent memory folder. A small extension can implement Agent Memory without changing Pi itself.
Setup guides
Harness-specific setup guides and copyable prompts are coming next. Where a harness does not support the complete format natively, the guide will show how to add as much support as the harness allows using extensions, hooks, or existing configuration.