API Overview

Everything you can do

Each row links to a reference page carrying the Python, Node.js and cURL form of the call, its request and response fields, its errors, and its scoping rules. Nothing is listed here that is not documented.

Memory

The data plane. Authenticated with an API key, and user-scoped through X-End-User-ID where the operation documents it.

Start at the Memory API overview if you are choosing between them.

Retrieval

Five ways to get memories back, and one to rebuild the vectors behind them. They differ in how much work the server does before answering.

Intelligence

Derived views over what is already stored. These read; they do not create memories.

Entities, tuning and audit

The vocabulary memories are classified against, the record of how one changed, and the signal that moves retrieval ranking.

Tenant data plane

A separate /v1 surface for multi-tenant applications that hold their own users. Episodic ingestion and recall live here rather than on the operations above.

Connectors

Bring content in from an external source. The five shared pages cover any provider; the per-provider pages cover what only that provider has. Every connector page walks the same sequence: create the connection, authorise it, configure what to read, sync, then inspect what arrived.

Organizations, projects and access

The control plane. Authenticated with a bearer access token rather than an API key, and governed by capabilities as well as scopes.

Data operations

Long-running work over project data belongs to the operations plane rather than the organization-resource hierarchy.

Webhooks

Receive events instead of polling for them. Endpoints can be paused, inspected, rotated and resumed; deliveries are recorded, retryable and replayable.

Billing

The public billing contract is intentionally read-only. Read the current plan and allowance; subscription-provider workflows and deployment financial administration are not public SDK operations.

What is not here

Two things developers ask for that MemorySync does not currently publish, stated plainly so nobody goes looking:

• There is no MCP server. The API carries everything one would wrap, but no MCP surface is published, so none is documented.
• There is no CLI. Use the Python or Node.js SDK, or cURL.

Request anatomy

Control-plane client contract

Make a first request

import os
from memorysync import MemorySyncClient

client = MemorySyncClient(
    api_key=os.environ["MEMORYSYNC_API_KEY"],
    base_url="https://api.memorysync.io",
    project_id=os.environ["MEMORYSYNC_PROJECT_ID"],
    end_user_id="usr_7f3a9c2e",
)

result = client.add(
    "The user prefers concise answers.",
    tags=["preference"],
)
print(result)

Read the result

{"id":101,"text":"The user prefers concise answers.","tags":["preference"],"is_summary":false,"created_at":"2026-05-04T12:30:11Z"}

Use the API safely

Choose the right client

Understand successful responses

Production integration checklist

• Create the client once in trusted server code and load credentials from a secret manager.
• Derive project and end-user scope from authenticated application state.
• Validate operation inputs before sending them and branch on the documented success variants.
• Treat 4xx responses as requests to correct; treat uncertain write failures as outcomes to reconcile.
• Log safe request identifiers and status information without logging API keys or sensitive memory text.
• Keep SDKs and generated types updated, while tolerating additive response fields.

Related reference