Agent Memory MCP

Authentication

MCP and REST share the same dashboard API keys (inf_ prefix). Team-scoped keys are required; local .env placeholder keys cannot be used for production MCP.

Transport

• Protocol: MCP Streamable HTTP (2025-11-25)
• Mode: stateless + enableJsonResponse (one POST → one JSON body; GET SSE listen disabled by default)
• mcp.json still needs Accept: application/json, text/event-stream (spec; missing values often return HTTP 406)
• Standard MCP methods supported: initialize, tools/list, tools/call, etc.

Recommended workflow

1. Call memory_list_libraries first (use description / writableTargets); optionally memory_suggest_write_library
2. Use memory_retrieve (combined layers) or memory_search / memory_recall / memory_wake_up for retrieval; pass allAccessible: true for every listable library
3. Proactively call memory_save_atom when the session reveals durable knowledge (decisions, conventions, preferences, root-cause fixes); search first to dedupe. Cross-team targets need sourceMemoryLibraryId + write grant. Multiple distinct facts in one turn: separate save per fact; parallel tool calls allowed
4. Call memory_correct_atom when search hits are outdated (do not save duplicates for the same fact). Multiple outdated atoms in one turn: one correct per atom; parallel tool calls allowed
5. Write tools default to async (202 + job.id): fast ack, background persist; do not expect immediate search hits
6. Writes require memory:write; MCP instructions inject SAVE MEMORY POLICY by default
7. All tools except list_libraries / suggest_write_library require memoryLibraryId or memoryLibraryLabel in arguments

Tools

Cursor configuration example

{
  "mcpServers": {
    "engra-memory": {
      "url": "https://mcp.engra.ai/api/v1/memory/mcp",
      "headers": {
        "Authorization": "Bearer inf_xxxxxxxx",
        "Accept": "application/json, text/event-stream",
        "Content-Type": "application/json"
      }
    }
  }
}

MCP Agent policy (server instructions)

Server instructions in the MCP initialize response guide when to retrieve, save, and correct. Configure in the console (soft constraints — models are not guaranteed to follow):

Wake policy mcp_wake_policy

off (agent decides) | session (memory_wake_up before first reply) | every_turn (search/wake_up before each user message). Per-library override available.

Save policy mcp_save_policy

off | proactive (default — proactive memory_save_atom; multiple distinct facts in one turn: separate save per fact, parallel tool calls allowed) | session_end (also require a structured summary atom before session ends).

Auto-correct autoCorrectEnabled

Per-library Settings toggle; when on with library-bound MCP, instructions inject CORRECT MEMORY POLICY so agents call memory_correct_atom instead of duplicate saves; multiple outdated atoms in one turn: one correct per atom, parallel tool calls allowed.

Default library / key override

Team Memory Settings or API key mcp_default_library_id; library-bound endpoints implicitly fix the library.

Custom instructions mcp_custom_instructions

Team-level append text (≤2000 chars) added to server instructions.

Configure at: each library → Settings → MCP Agent policy; team Memory Settings → MCP default library; /dashboard/keys → per-key default library.

HTTP integration

Custom MCP clients should follow the Streamable HTTP spec: POST requests need Accept including application/json and text/event-stream. This server primarily responds with application/json (JSON-only mode does not offer GET SSE by default). Cursor and similar clients should set Accept and Content-Type in mcp.json headers; otherwise HTTP 406 is common. Full example: skills/engra-memory/mcp.example.json.