Synap Memory

synap-mcp-server

Hosted remote MCP server (Streamable HTTP) that re-fronts the existing Synap public REST operations as MCP tools, so no-code platforms (Gumloop, n8n) can give their agents persistent memory with nothing but a pasted MCP URL and a Bearer token.

It is a stateless adapter — no new backend, storage, or pipeline. Each tool call maps to one existing REST operation, and the incoming Bearer synap_ token is forwarded verbatim to synap-cloud, which owns auth.

• Public endpoint (prod): https://synap-mcp.maximem.ai/mcp
• Health: https://synap-mcp.maximem.ai/health

Tools

Scoping: one contract, two modes

The REST API enforces this. It is not advice, and it has been live on production since 2026-08-26 07:07 UTC.

GET /api/v1/auth/whoami reports the mode as user_context_isolation. This server reads it once per token and refuses a customer_id locally on a B2C instance, so a model gets a sentence telling it what to send instead of a status code it cannot act on. If the mode cannot be read, nothing is refused and the API stays authoritative.

Pass no IDs at all ⇒ client scope (shared per credential), which is valid in both modes. Note: client-scope writes are not surfaced on the dashboard Memories page (relational), so per-user scoping is recommended when dashboard visibility matters. Fill user_id from an n8n expression / Gumloop input, and pass the same value on log_exchange and recall_context so writes and reads address the same person.

Host header. The server disables the MCP transport's DNS-rebinding check (it sits behind a proxy + Bearer auth), so any reverse proxy can forward the real Host — no Host-rewrite hack needed.

Run locally

pip install -e ".[dev]"

SYNAP_API_URL=http://localhost:8000 uvicorn synap_mcp_server.server:app --port 8090 curl http://localhost:8090/health # {"status":"ok",...}

Reporting the turn to the anticipation agent

An MCP server is not an agent. It exposes tools to somebody else's agent and is handed one stateless HTTP call at a time, so most of what an agent does is invisible here: the caller's tool calls, its tool results and its reasoning never reach this process, and nothing tells a stateless server that a conversation has ended. Those are not reported, because there is nothing to report.

The turn does exist. log_exchange is handed the user's message and the assistant's reply as arguments. Set MCP_STREAM_EVENTS=true and each logged exchange is also reported as a conversation turn on POST /v1/events/batch, the HTTP door onto the same listening path the gRPC stream feeds, so the anticipation agent sees the turn and not only the memory extracted from it. This server does not use the SDK's gRPC stream, as the other integrations do: it holds no SDK, keeps no state between requests, and every request can carry a different tenant's key, so a stream per key would be an unbounded pool of connections that nothing closes. Moving it onto the SDK would be a rewrite of its transport for a caller that has no use for what the stream brings back.

Nor does it replace anything of Synap's compaction: the caller's agent keeps its own history and no tool here returns a conversation.

It is off by default, on purpose. Turning it on adds a second backend call per log_exchange, writes the turn to conversation history in addition to the long-range document, and runs the anticipation agent, which bills model calls.

It is not on by default for a second reason. Synap extracts from a conversation when it is compacted, which starts once ten messages have been recorded. Recording every exchange as turns *instead of* the long-range document would hold the memory back until then (the next recall_context finds nothing, and there is no ingestion_id to check); recording them *as well* extracts each exchange twice.

What does not change either way: log_exchange still queues the long-range document and still returns the ingestion_id that check_memory_status and wait_for_processing are built on. A turn report is best effort and can never alter what a tool answers. A turn with no user_id or no conversation_id is not reported at all, because the events route requires both.

Test

pip install -e ".[dev]"

pytest -q

Config (env)

No Synap API key is configured on the server — the end user's key arrives per-request as a Bearer token.

See the Synap MCP documentation for platform setup guides (Gumloop, n8n) and deployment details.