How is Mnemosyne different from my bot’s built-in memory?

Built-in memory belongs to one bot, so it does not travel when you run a second bot or rebuild the first. Mnemosyne is a memory service you host, so several MCP clients can read and write the same store.

What endpoint should I enter?

Mnemosyne’s native transport is SSE, served at /sse, so enter that HTTPS address including the /sse path — the connector recognises it and speaks SSE. A Streamable HTTP gateway in front of the server also works.

Where is the data stored?

In a single SQLite file on the server you run, which makes it easy to back up, inspect and move. Recall combines vector similarity, full-text search and an importance weighting.

← Home

Guide

Mnemosyne

Tell one bot how you like things done and the other one still does not know. Mnemosyne is a memory that lives outside any single agent: durable memories in one SQLite file, retrieved by a mix of meaning, wording and importance, and readable by every MCP client you point at it. Works with OpenClaw and Hermes.

Why a Memory Outside the Bot

Each framework already remembers things — OpenClaw and Hermes both do. That memory belongs to that bot. Run two, or rebuild one, and the knowledge does not travel.

Mnemosyne inverts the arrangement. The memory is a service you own, and the agents are clients of it. A preference recorded through one bot is available to the next, and to a coding agent on your laptop pointed at the same server.

It is MIT licensed and stores everything in a single SQLite file, which is the detail that makes it comfortable: you can back it up, copy it, inspect it with ordinary tools, and move it somewhere else without an export process.

How It Retrieves

Recall is hybrid rather than purely semantic: vector similarity for meaning, full-text search for exact wording, and an importance weighting so the things that matter surface first. In practice that combination is what stops a memory store from either missing the paraphrase or burying the one line you actually needed under twenty near-matches.

Connect It

Run Mnemosyne and start its MCP server. Its native transport is SSE — mnemosyne mcp --transport sse — served at /sse.
Put it behind HTTPS so your instance can reach it, and put authentication in front of it. A memory server is a list of things you told your agents in confidence; it does not belong open on the internet.
Open the dashboard, go to Tools, pick a running bot, find Mnemosyne, and enter your endpoint — including the /sse path — plus the authentication you configured.

Giving the /sse address directly is the right thing to do rather than a workaround: the connector recognises that path and speaks the SSE transport to it. If you already front the server with a Streamable HTTP gateway, that address works too.

What to Ask For

“Remember that I want short answers with the reasoning left out.”
“Remember this decision and why we made it — it matters.”
“What do you already know about how I like reports written?”
“Recall anything relevant before you answer this one.”

Under the hood those are a remember call, which takes the content along with a scope and an importance, and a recall call that takes a query and a limit. You do not write them yourself — but knowing importance is a real field explains why saying “this matters” genuinely changes what comes back later.

Which Memory for What

Mnemosyne — shared, durable, yours. Facts and preferences that should outlive any one bot.
The framework’s own memory — per-bot conversational memory, with no setup at all.
Codebase Memory — structural knowledge of a repository, which is a different kind of thing again.
Skills — procedures you want followed, rather than facts to recall.

Troubleshooting

The connection fails. Check the endpoint includes the transport path your server actually serves, and that it is reachable over HTTPS from outside your network.
Nothing is recalled. An empty store recalls nothing. Ask the bot to remember a few things first, then ask about them.
The wrong memories come back. Ask more specifically, and mark the things that matter as important when you save them — ranking is part of how recall works here.
Two bots disagree. Confirm both are pointed at the same server. Sharing is the feature, but only if they share the same endpoint.

Try It

Connect it to one bot, tell it two or three durable things about how you work, then connect a second bot to the same server and ask it what it knows about you. That second answer — from a bot you never told — is the whole idea.