Frequently asked questions

Mnemoverse is a persistent memory API for AI agents. You store preferences, decisions, lessons, and context once, and recall them from any tool (Claude Code, Cursor, VS Code, Windsurf, ChatGPT, Python, or any HTTP client) with a single API key or OAuth.

A vector database does similarity search over stored embeddings. Nothing learns or changes. Mnemoverse is a memory: it forms associations between concepts (Hebbian), tunes recall from outcome feedback (Rescorla-Wagner), consolidates and forgets over time (HDBSCAN clustering with Von Restorff protection for distinctive items), and expands queries automatically. It improves with use.

They are complementary. RAG answers "what do the documents say?" by retrieving chunks of a fixed corpus. Mnemoverse answers "what did we discuss, decide, or learn last time?" It accumulates user- and project-specific experience across sessions. Many teams use both.

The client libraries (Python SDK, MCP server) are open source under MIT. The memory engine itself is a hosted service. You can self-host on Enterprise plans.

Mnemoverse works across Claude Code, Claude Desktop, Cursor, VS Code, Windsurf, ChatGPT, the Python SDK, and any HTTP client with one API key or OAuth. The same memories are available everywhere.

One command: claude mcp add mnemoverse -s user -e MNEMOVERSE_API_KEY=mk_live_YOUR_KEY -- npx -y @mnemoverse/mcp-memory-server@latest. The MCP server exposes 10 memory tools to the agent.

Yes, through the hosted MCP connector or the REST API directly. Custom GPTs you already built keep working; as of August 2026, OpenAI limits creating new ones to Business, Enterprise and Edu workspaces — their rules here change often, so check their help center for the current state. The connector and the REST API work on every plan.

Yes: pip install mnemoverse (live on PyPI). It wraps write, read, feedback, stats, and consolidation with typed errors.

@mnemoverse/mcp-memory-server is an open-source Model Context Protocol server (listed on the MCP Registry) that gives any MCP-capable agent 10 memory tools over a standard stdio transport, the same set as the hosted connector.

Yes. Free is $0: 1,000 queries/day, 10,000 atoms, 60 requests/minute, no credit card required.

Any read or write call to the memory API. Daily limits reset every 24 hours; rate limits are per minute. Paid plans raise all three limits.

Yes. Plans are month-to-month (Pro $29/mo, Team $149/mo) and you can change or cancel at any time from the console.

Custom limits, dedicated infrastructure, SSO/SAML, audit logs, data-residency options, self-hosting, and a custom SLA. Contact sales to scope it.

Mnemoverse keeps every tenant's memories under a separate org identifier; cross-tenant reads return no results. Requests are authenticated per API key.

No. Mnemoverse does not train models on your stored memories.

Keys are hashed with SHA-256 before storage and compared in constant time to prevent timing attacks. The raw key is shown once at creation.

On managed PostgreSQL 17 with pgvector (US region today). Enterprise plans can choose data residency and self-hosting.

Mnemoverse strengthens Hebbian associations between concepts recalled together, so related memories surface together over time. Outcome feedback (a +1/-1 signal when a recalled memory was useful) tunes future recall via a Rescorla-Wagner update.

Yes, and it's the one "bug" Mnemoverse engineered deliberately and carefully. A consolidation step (HDBSCAN clustering) merges redundant memories, while Von Restorff protection keeps distinctive, singleton items from being lost. Engineered forgetting is what makes persistent memory usable: without it the store just grows into noise. And forgotten doesn't mean gone: the underlying memories are still retained, so a consolidated item can be retrieved if it's ever needed; we simply don't expose that path by default, to keep recall sharp.

Yes. The underlying SLoD approach is published (arXiv:2603.08965) and was peer-reviewed and accepted at the GRAAI workshop (IEEE WCCI 2026). Mnemoverse is benchmarked on public memory benchmarks including LoCoMo and LongMemEval.