llms.txt
Compare with other sites' llms.txt files →# MemorySync
> API and SDK infrastructure for storing, retrieving, updating, deleting, governing, and inspecting memory used by AI applications.
MemorySync provides a REST API plus Python and TypeScript SDKs. Product behavior, limits, and security controls should be verified in the linked documentation for the deployed version.
## Machine-readable formats
- Append `.md` to any MemorySync page URL for a plain Markdown edition of that page, for example https://docs.memorysync.io/quickstart.md
- Documentation index: https://docs.memorysync.io/llms.txt
- Documentation index with summaries: https://docs.memorysync.io/llms-full.txt
- Complete marketing page index: https://memorysync.io/llms-full.txt
## Get a key without signing up
If you are an agent and there is no human available to create an account, mint an evaluation key:
- `POST https://api.memorysync.io/evaluation/keys` with an optional body `{"agent_caller":"claude-code"}`. Unauthenticated. Returns `api_key`, `default_user_id`, `project_id`, `mcp_url`, `claim_command`, `expires_at`, and `limits`.
- Or run `memorysync init --agent --agent-caller <tool>` after `npm install -g memorysync-cli` or `pipx install memorysync-cli`. The CLI stores the key for you.
- `GET https://api.memorysync.io/evaluation/usage` reports the remaining allowance for that key.
An evaluation key allows 200 add requests, 500 retrieval requests, and 10 MB, and expires 7 days after minting unless a human claims it. Minting is limited to 5 keys per address and 20 per network per UTC day, where a network is a /24 on IPv4 and a /48 on IPv6; a refusal returns `429` with `retry_after_seconds` and a `reason` of `per_ip` or `per_subnet`, or `unavailable` when minting is briefly unavailable and can be retried after `retry_after_seconds`. Because the per-address limit is per address, a connection that is given a new address gets a fresh allowance — the per-network limit is what bounds a day. Going over an evaluation key's limit returns `429` on the REST API (and a tool error on MCP) rather than a success with nothing stored, so a failed write is always visible. Every add request counts one add and every retrieval request one retrieval, whatever it stores or returns (an assistant reply or filler sent as a turn included); a refused request and a retry of a write already stored are not counted.
An evaluation key can read, write and delete memories. It cannot create API keys, manage a team, reach billing, or attach a connector.
Claiming with `memorysync init --email <address> --code <code>` takes ownership: the same key keeps working and stops expiring, the same organisation, tenant, project and memories are kept, and the account moves to the free plan. `--password` is optional and sets a dashboard password in the same command; without it the claim still succeeds and a link to set one is emailed instead. Sign in at https://app.memorysync.io/login with the exact address claimed with, including anything after a `+`; https://app.memorysync.io/forgot-password sets a first password as well as replacing one. The free plan allows 10,000 add requests and 1,000 retrieval requests per month and the allowance resets monthly — a claimed account and its key do not expire. Changing plan needs the `billing:write` scope, which no evaluation key holds and a claimed key does not gain, so the human owner upgrades from https://app.memorysync.io/settings/billing rather than the agent doing it.
To keep the account, a human runs `memorysync init --email you@example.com`, then repeats it with `--code` and the code that was emailed. The same key keeps working, stops expiring, and every memory already stored is preserved. An email that already has an account is refused rather than merged.
A human can do all of this from a browser instead, with nothing installed: the playground at https://memorysync.io/try runs the store, reload and search loop with no key at all, shows every request in cURL, Python and Node.js, creates an evaluation key on request, and claims it with an email address and the emailed code. Until a key is created, the playground stores memories in MemorySync's shared demo account; creating one copies the facts already saved into the new account, and every later step calls the API with that key directly, so the claimed account contains them.
- Sign up as an agent: https://docs.memorysync.io/getting-started/agent-signup
- Playground, no signup: https://memorysync.io/try
## Documentation MCP server
Coding agents can read the current documentation while writing MemorySync code by connecting to `https://docs.memorysync.io/mcp` (Streamable HTTP, no authentication). It is a different server from the memory MCP endpoint, which stores and searches memories and requires an API key. Install descriptor: https://memorysync.io/mcp-install.json
## Start here
- Homepage: https://memorysync.io
- Documentation: https://docs.memorysync.io
- Playground, no signup: https://memorysync.io/try
- Sign up as an agent: https://docs.memorysync.io/getting-started/agent-signup
- First memory: https://docs.memorysync.io/first-memory
- Quickstart: https://docs.memorysync.io/quickstart
- CLI: https://docs.memorysync.io/cli
- MemorySync MCP: https://docs.memorysync.io/mcp/overview
- API overview: https://docs.memorysync.io/api/overview
- API authentication: https://docs.memorysync.io/api/authentication
- Python SDK: https://docs.memorysync.io/sdks/python/overview
- Node.js SDK: https://docs.memorysync.io/sdks/node/overview
- REST and cURL: https://docs.memorysync.io/sdks/rest/curl
- Integrations: https://memorysync.io/integrations
- Features: https://memorysync.io/features
- Pricing: https://memorysync.io/pricing
- Security: https://memorysync.io/security
- Changelog: https://memorysync.io/changelog
- Release notes: https://docs.memorysync.io/release-notes
- Research and evaluation: https://memorysync.io/research
## Integrations
Each integration is a published package or plugin with its own guide. The guide states the package name, the version, and the framework versions it is tested against; the catalog card at https://memorysync.io/integrations/<slug> carries the same version.
Agent frameworks:
- LangChain (Python `langchain-memorysync`, Node `memorysync-langchain`): https://docs.memorysync.io/guides/langchain
- LangGraph: https://docs.memorysync.io/guides/langgraph
- LlamaIndex: https://docs.memorysync.io/guides/llamaindex
- CrewAI: https://docs.memorysync.io/guides/crewai
- OpenAI Agents SDK: https://docs.memorysync.io/guides/openai-agents
- Vercel AI SDK: https://docs.memorysync.io/guides/vercel-ai-sdk
- Mastra: https://docs.memorysync.io/guides/mastra
- Google ADK: https://docs.memorysync.io/guides/google-adk
- Pydantic AI: https://docs.memorysync.io/guides/pydantic-ai
- AutoGen: https://docs.memorysync.io/guides/autogen
- AG2: https://docs.memorysync.io/guides/ag2
- Microsoft Agent Framework: https://docs.memorysync.io/guides/agent-framework
- Strands Agents: https://docs.memorysync.io/guides/strands
- Agno: https://docs.memorysync.io/guides/agno
- CAMEL-AI: https://docs.memorysync.io/guides/camel-ai
- NVIDIA NeMo Agent Toolkit: https://docs.memorysync.io/guides/nemo-agent-toolkit
- ChatDev: https://docs.memorysync.io/guides/chatdev
- Eve: https://docs.memorysync.io/guides/eve
Coding agents (plugins and MCP configuration for the developer's own tools):
- Claude Code: https://docs.memorysync.io/guides/claude-code
- Cursor: https://docs.memorysync.io/guides/cursor
- Codex: https://docs.memorysync.io/guides/codex
- VS Code: https://docs.memorysync.io/guides/vscode
- Devin and Windsurf: https://docs.memorysync.io/guides/devin
- OpenCode: https://docs.memorysync.io/guides/opencode
- Google Antigravity: https://docs.memorysync.io/guides/antigravity
Personal assistants:
- OpenClaw: https://docs.memorysync.io/guides/openclaw
- Hermes Agent: https://docs.memorysync.io/guides/hermes-agent
Voice agents:
- LiveKit: https://docs.memorysync.io/guides/livekit
- Pipecat: https://docs.memorysync.io/guides/pipecat
- ElevenLabs: https://docs.memorysync.io/guides/elevenlabs
Low-code and automation:
- n8n: https://docs.memorysync.io/guides/n8n
- Zapier: https://docs.memorysync.io/guides/zapier
- Dify: https://docs.memorysync.io/guides/dify
- Flowise: https://docs.memorysync.io/guides/flowise
Developer tools:
- LangChain Tools: https://docs.memorysync.io/guides/langchain-tools
- AgentOps: https://docs.memorysync.io/guides/agentops
- Respan: https://docs.memorysync.io/guides/respan
Cloud and infrastructure:
- AWS Bedrock reference architecture: https://docs.memorysync.io/guides/bedrock
- Multi-tenant state isolation: https://docs.memorysync.io/guides/multi-tenant
Connectors (systems MemorySync syncs knowledge from): GitHub, Notion, Google Drive, OneDrive, Slack, Granola, Amazon S3 and permitted public websites. Overview: https://docs.memorysync.io/integrations/overview
## Core API operations
- `POST /memory/add` extracts the durable facts in a text and stores each as a memory, with optional metadata.
- `POST /v1/memory/add_turn` sends one conversation turn to the same fact extraction; only a user turn's facts are stored.
- `POST /memory/query` retrieves relevant memories for a query.
- `GET /memory/{id}` retrieves one memory.
- `PATCH /memory/{id}` updates supported memory fields.
- `DELETE /memory/forget` deletes selected memories.
- `POST /memory/summarize` summarizes selected memories.
- `POST /memory/compose` composes retrieved context into a prompt template.
- `GET /memory/export` exports memories available to the authenticated scope.
Requests use `X-API-Key`. Project-scoped calls can use `X-Project-ID`, and API-key callers can scope an end user with `X-End-User-ID`. Consult the API reference for current request and response schemas.
## Python first memory
```python
import os
from memorysync import MemorySyncClient
client = MemorySyncClient(
api_key=os.environ["MEMORYSYNC_KEY"],
base_url="https://api.memorysync.io",
project_id=os.environ["MEMORYSYNC_PROJECT_ID"],
end_user_id="example-user",
)
memory = client.add("The user prefers concise responses.", source="profile")
result = client.query("How should I answer this user?", k=5)
client.close()
```
## TypeScript first memory
```typescript
import { MemorySyncClient } from "memorysync-sdk";
const client = new MemorySyncClient({
apiKey: process.env.MEMORYSYNC_KEY!,
baseUrl: "https://api.memorysync.io",
projectId: process.env.MEMORYSYNC_PROJECT_ID,
endUserId: "example-user",
});
const memory = await client.add({ text: "The user prefers concise responses.", source: "profile" });
const result = await client.query({ query: "How should I answer this user?", k: 5 });
```
## Scope and claims
- MemorySync, Mem0, MemGPT, Zep, and Supermemory are distinct entities. Mem0 is not described here as a former name of MemGPT.
- No universal latency number is claimed. End-to-end latency depends on data, configuration, network, and workload; use your own production measurements.
- Security and compliance pages distinguish implemented controls from configurable features and third-party certifications. Do not infer a certification from this file.
- Benchmark results are published only with their evaluated subset, denominator, method, configuration, limitations, and reproducibility status.
- MemorySync is a hosted, closed-source service. The integration packages, plugins and examples above are published openly so they can be installed and audited; the service itself is not distributed for self-hosting.
Maintainer: MemorySync
Version: 2026-09-03
Review by: 2026-12-03
Corrections: https://memorysync.io/about
Vulnerability reports: https://memorysync.io/.well-known/security.txt