Mantle
The self-hosted AI brain that remembers your whole life
Start new thread

Mantle - The self-hosted AI brain that remembers your whole life

by•
A self-hosted AI brain for your whole life. Feed it notes, email, files, calendar and documents. It extracts summaries, facts and an entity graph as things arrive, so search actually finds what you meant. Ask the built-in assistant anything, or let specialist agents research, write, query tables and run jobs in the background. Everything lives on your own server. Open source, no vendor, and your data never leaves the machine.

Add a comment

Replies

Best
Maker
📌
Hi Product Hunt 👋 I built Mantle because my own memory kept living in other people's databases. Notes in one app, email in another, files somewhere else, and an assistant that could only ever see whichever window happened to be open. Mantle is the opposite. It runs on my own server, ingests everything I feed it, and understands it as it arrives: a summary, the extracted facts, entities linked into a graph, passage-level embeddings. So when I ask "which revision of that document is current, and what changed", it answers from the whole brain instead of grepping for keywords. Four things I care about in it: It is awake. Heartbeats run agent routines on a schedule, so it reads the inbox, files things and sends me a morning briefing without me ever opening the app. A brain that only answers when asked is a database with a chat skin. There is no "new chat". Recent turns stay raw, older conversation is compressed into topic digests, durable facts are kept current and superseded rather than duplicated, and a recall specialist can replay the actual words of any past conversation. It picks up where we left off, last week or last year. One Postgres, no zoo. Vector search, the knowledge graph, full-text, job queues, realtime and auth all live in one database. No Pinecone, no Neo4j, no Redis. That is why the whole system idles at about 2.5 GB of RAM and a 4 GB VPS carries an entire brain. You choose the models. Frontier models where quality matters, cheap ones for background compression, and if you want nothing to leave the box there is a bundled local embedder and it will talk to models on your own tailnet. It has been my daily driver since June, on my real inbox and my real files, which is also why it has opinions about things like knowing when a document has been superseded. Two questions I would genuinely like answered: what would you feed a brain like this first? And if you have tried self-hosting an AI setup before, what made you give up on it? Happy to go deep on the architecture, the memory model, the prompt-injection boundary or the deployment.