OpenCode memory: AGENTS.md, then a clean slate
OpenCode is fast, scriptable, and honest about state: your project instructions are files, and the rest of the session is a session.
SST. Sources checked 2026-09-25.
How OpenCode handles memory today
OpenCode uses AGENTS.md for project rules, with /init to scan a repository and generate one. A global AGENTS.md in the config directory covers all sessions, and files are discovered from the current directory downward.
OpenCode supports MCP servers through its configuration, which is how a memory service attaches to the agent loop.
What OpenCode forgets
A stateless agent with good instructions still starts at zero on history.
- Session context at exit, including decisions that never became code.
- Cross-project context, since instructions and history are directory-scoped.
- The record of what was tried and rejected, which is exactly what a new session needs.
- Anything from your other tools.
The fix with MemoryRouter
Register MemoryRouter as an MCP server in the OpenCode config and the agent can recall decisions and store new ones in your vault.
- 1
Add the server to your OpenCode config
opencode.json
{ "mcp": { "memoryrouter": { "type": "remote", "url": "https://mcp.memoryrouter.ai/mcp" } } } - 2
Authenticate on first use
Complete the OAuth flow and choose the vault. Credentials stay with the client, not in the config file.
- 3
Ask for recall explicitly
Prompt the agent to check memory when you start work on a module you have touched before. Tool selection is model-directed.
Connect OpenCode over MCP and share the vault with Claude Code and Codex. 14-day free trial.
Related
OpenCode memory FAQ
Sources
Sources checked 2026-09-25. Tool behavior changes; check the vendor docs before relying on a detail.