Memstate + OpenCode
Give OpenCode persistent, structured memory via MCP.
OpenCode is an open-source AI coding agent built by the SST team. It supports MCP natively via its config file, making it straightforward to add Memstate as a persistent memory layer that survives across sessions.
Get Your API Key
Sign up for free — no credit card required.
Sign Up FreeAlready have an account? Go to API Keys →
Add MCP Configuration
Add this to your OpenCode MCP config file. Replace YOUR_API_KEY_HERE with your key.
{
"$schema": "https://opencode.ai/config.json",
"mcp": {
"memstate": {
"type": "local",
"command": ["npx", "-y", "@memstate/mcp"],
"enabled": true,
"environment": {
"MEMSTATE_API_KEY": "YOUR_API_KEY_HERE"
}
}
}
}Save to:
./opencode.jsonc (project root)Config file location
OpenCode reads MCP config from opencode.jsonc in your project root, or from the global config at:
- macOS / Linux:
~/.config/opencode/config.jsonc - Windows:
%APPDATA%\opencode\config.jsonc
The MEMSTATE_API_KEY is passed as an environment variable to the MCP server process.
Add Agent Rule
This rule tells OpenCode to load your project context before tasks and save summaries after — automatically, every session.
# Memstate MCP - Memory Storage
This project uses Memstate MCP for persistent, versioned memory across sessions.
## REQUIRED: Start and end every task with memory
**BEFORE starting any task**, load existing context so you don't redo past work or revert intentional decisions:
```
memstate_get(project_id="<your_project>")
```
**AFTER completing any task**, save what you did so the next session has context:
```
memstate_remember(project_id="<your_project>", content="## Task Summary\n- What was done\n- Key decisions made\n- Files modified", source="agent")
```
These two calls are mandatory. Skipping them means the next session starts blind.
## Tool reference
| Tool | When to use |
|------|-------------|
| memstate_get | **Start of every task.** Browse project tree or fetch content at a keypath. |
| memstate_remember | **End of every task.** Save markdown summaries, notes, decisions. |
| memstate_search | Find memories by meaning when you don't know the exact keypath. |
| memstate_set | Store a single key=value fact (e.g. config.port). Not for summaries. |
| memstate_history | View version history of a keypath. |
| memstate_delete | Soft-delete a keypath (history preserved). |
## Project naming
Use a short snake_case name matching your repo or topic (e.g. my_app, api_service). All related memories should share the same project_id.Create this file at: opencode.jsonc → rules field — OpenCode rules docs
Test It — Onboard Your Project
Restart OpenCode, open a project, and paste this prompt. It will create your first memories and confirm everything is working.
Replace <your_project> with a short name for your repo (e.g. my_app).
I'm onboarding Memstate AI memory for this project. Please:
1. Analyze this codebase and write a concise high-level architecture overview in markdown — covering the main components, tech stack, key directories, and how data flows through the system.
2. Save it to Memstate using: memstate_remember(project_id="<your_project>", content="<the markdown>", source="agent")
3. Then call memstate_get(project_id="<your_project>") and show me the memory tree so I can confirm it worked.What to expect
Your agent will analyze the codebase, write an architecture overview, save it to Memstate, then display the memory tree. You'll see structured memories like project.my_app.architecture — this context is now available in every future session automatically.
Related Resources
Mem0 vs Memstate
See how Memstate compares to Mem0 on the LoCoMo benchmark. Memstate scores 5.3x higher on fact recall accuracy (92.2% vs 17.5%) and 4.7x better on conflict detection (95.0% vs 20.2%).
What is AI Agent Memory?
Learn how structured, versioned memory differs from traditional RAG and vector search.