Memory MCP: five-minute Cursor setup
Streamable HTTP MCP tools for libraries, search, recall, wake-up, and atom writes—plus SKILL.md workflows for production agents.
What you get
Engra exposes Agent Memory over MCP at /api/v1/memory/mcp. Transport is Streamable HTTP—no local stdio bridge, no extra daemon on the developer machine.
Published tools include: • memory_list_libraries — enumerate libraries (memory:read) • memory_search — L3 semantic search (memory:read) • memory_recall — L2 topic recall (memory:read) • memory_wake_up — L0+L1 session context (memory:read) • memory_save_atom — persist a new atom (memory:write; default async 202) • memory_correct_atom — correct an existing atom (memory:write; default async 202) • memory_list_atoms — paginated atom listing (memory:read)
Pair MCP with Agent Skills: SKILL.md files document when to wake up memory, how to search, and how to save durable facts back into the atom graph.
Prerequisites
You need an Engra account, a Memory library, and a platform API key with memory:read (and memory:write if you will save atoms). Create keys in the console under API keys.
Cursor (or any MCP client supporting Streamable HTTP) must be able to reach your Engra origin with the key in Authorization or X-API-Key headers as documented.
Minimal integration path
1. Copy the MCP endpoint URL from the docs. 2. Add the server in Cursor MCP settings with your API key. 3. Install the published Skill so the agent knows when to call wake_up vs search. 4. Start a new chat and confirm memory_list_libraries returns your library.
Recommended session flow: memory_wake_up at session start → memory_search before tool-heavy turns → memory_save_atom when the user states a durable preference or decision. Write tools return quickly (202 + job id) so IDE chats stay responsive; indexing finishes in the background.
Common pitfalls
Skipping wake_up means the agent relies only on L3 search and misses L0 identity + L1 narrative context. Always wake at session start for multi-turn coding agents.
Writing atoms without Scope/Topic metadata dilutes retrieval. Follow the Skill playbook: attach scope, topic, and memoryLibraryId when saving (legacy write field persona still accepted).
When to use REST instead
Custom backends, batch ingestion, and multi-tenant orchestration should call /api/v1/memory directly. MCP is optimized for IDE and chat agents; REST is optimized for services you control.