The core idea that sets this project apart is the addition of inline ACLs on the prose of the wiki pages; the inspiration came to me, in all honesty, when social media was full of the redacted documents from the Epstein files; the wikis are written like that, every paragraph, every item of a list, every sensitive piece of data has its own ACL, which is applied already during the autocapture of messages and at the ingest of documents and media. In practice: my partner can tell the assistant something and our kid can't get it out of it, "we're out of milk" belongs to the whole household and not to me; the same page is served differently to each reader, redacted by the engine before any text reaches the agent. Memory is organized into wikis, every user, group and agent has their own wiki and other wikis can emerge autonomously.
To reach this goal the server uses various components that use configurable LLMs, from the simple dedup, which you can easily run locally, to the reconciler that works during the REM phase (yes, there's a light dream and a REM dream) and reorders the pages that during the day, for efficiency, may have been written roughly. Engineering wise it's a single Rust binary, MCP endpoint (Streamable HTTP, with OAuth) and admin dashboard on the same port.
Recall is just as important: an initial RAG pass finds the entry doors into the various wikis, then the "navigator" follows the links of each page that fit the purpose of the search, to get the whole context of the data being looked up or the topic being discussed; this funnel navigation itself creates the context, ready to hand to the connected agent, and since it follows text written as prose it also surfaces stray memories, that is, memories nobody asked for but that can be important in the context being discussed, for example a user who once specified a food intolerance, when the agent is asked for a recipe for the group that user belongs to. The prose isn't there for beauty or for humans to read, the prose links the facts, so that no fact is disconnected from the rest but everything is part of one huge narrative, which is however made extremely efficient by the metadata.
These are the peculiarities, for the rest I took the best of what I found in the other existing memories and integrated it. Now I've realized that this memory system, initially born for a family, can be great for a team too and, with some future adjustments, even for small and medium businesses.
Honest limits: the internal LLM needs a reasonably capable model (small local ones route poorly, it's documented), and Windows support is best-effort for now. The repo is mostly vibe-coded, I'm aware of it, but I've been a developer since the last millennium and I would never have managed to finish it in a reasonable time; anyway it's in production and I've been using it for months, fixing the problems as they come out. License: AGPL, self-hosting is free forever; the sustainability model is a commercial license for closed SaaS cases.
Now I'm simply looking for people who can put it to the test, maybe someone who has a team or wants to try it with their family, but it also works for a single user with several agents.
Built for my family, released for yours. https://github.com/Fr4nZ82/mwe-mcp