Connect your coding agent

Your coding agent can read from and write to RememberStack directly, as a set of MCP tools (Model Context Protocol, the standard way agents call external tools). Once connected, the agent can look up what the project decided, who owns what, and what changed, and it can store new notes as it works.

The remember package includes an MCP server that runs next to your agent and talks to your engine. remember setup writes the configuration for you:

pip install remember
remember setup --self-hosted

It detects which agents you use in the current directory and on your machine (Cursor, Claude Code, Claude Desktop, Codex, Antigravity) and configures each one. To configure one agent only:

remember setup --self-hosted --agent cursor   # or claude, codex, agy

Add --dry-run to see what it would write, and --api-url if your engine is not at http://localhost:8000. Then check the result:

remember doctor

The files remember setup writes, and how to write them by hand, are in the MCP reference.

For an agent that connects to a URL rather than starting a program, run the server over HTTP and point the agent at http://127.0.0.1:8765/mcp:

remember mcp --transport http --api-url http://localhost:8000
remember setup --mcp-url http://127.0.0.1:8765/mcp

Cursor, Claude Code and Codex get the URL; Claude Desktop and Antigravity still start remember mcp themselves.

The HTTP server holds no key: it passes each agent's Authorization header to the engine. See MCP tools: Over HTTP.

What the agent can do

The remember mcp server gives the agent these tools:

Tool	What it is for
`resolve_entity`	Find which person, system or thing a name refers to. Reports ambiguity instead of guessing.
`facts_context`	The current facts about the entities in a question, with evidence and time windows. The usual first call.
`claims_and_sources_context`	What sources said, with the passages. Use it when the exact wording matters.
`combined_context`	Facts and claims together, when the agent wants everything in one call.
`ingest`	Store a note, a decision or a file.
`pipeline_readiness`	Check whether stored documents have finished processing.
`delete_document`	Remove a document from the memory when the user asks.
`query_sql` and six related tools	Read-only SQL over the memory, for questions the operations above do not cover.

Run remember mcp --read-only instead of remember mcp to leave out the two tools that change memory (ingest and delete_document).

Teach the agent the order

Agents use memory best when they follow one order: resolve the entity, then ask for facts, then fall back to claims and sources, then to SQL. remember setup writes this guidance as a Cursor rule (.cursor/rules/remember.mdc) and an Antigravity skill (.agents/skills/remember/SKILL.md). For other agents, put this in the system prompt or project instructions:

You have a project memory available through the "remember" tools.
1. When a question names a person, system or thing, call resolve_entity first.
   If it returns more than one candidate, ask which one is meant.
2. Call facts_context for what is currently true. Pass time={"mode": "history"}
   when the question is about how something changed.
3. When the exact wording or the source matters, call claims_and_sources_context.
4. If the memory reports it knows nothing, say so. Do not guess.
5. Quote evidence with its document when you rely on it.

Give an agent context: which operation fits which question.
Handle unknowns and ambiguity.

Connect your coding agent

What the agent can do

Teach the agent the order

Next