320 hand-tuned extractors plus spaCy read every conversation for facts worth keeping, and keyword search and embeddings merge through Reciprocal Rank Fusion to pick the right ones next time. Close the chat — it still knows your name, your stack, and what you're in the middle of.
An OpenAI-compatible front door means it drops into Open WebUI, SillyTavern, n8n, or anything that speaks the API. Ollama, llama.cpp, vLLM, SGLang and LM Studio behind it; Claude, GPT-5, Gemini and OpenRouter alongside them if you turn cloud on.
The whole memory brain — extraction, retrieval, consolidation, the knowledge graph — runs on your hardware. Cloud providers are opt-in and off by default, and local_only mode means nothing leaves the machine at all.
Most AI chat apps have the same blind spot: every conversation starts from zero. You tell the model your name, your preferences, what you're working on — and an hour later, in a fresh chat, it's a stranger again. The models got extraordinary. The part that remembers you never got built.
The second problem shows up the moment more than one person uses the same box. Everyone's context crashes into one pool, and Sarah asks about her meeting notes and gets Mike's deploy script back. That isn't a memory problem, it's an isolation problem, and it needs solving at the same layer.
UPtrim solves both in the one place that can see everything without owning anything: the wire between the chat app and the model. It watches each conversation, extracts what's worth keeping, files it under the right person, and quietly puts the relevant pieces back into the next prompt. The frontend thinks it's talking to the LLM. The backend thinks it's getting an ordinary request. Nothing in your stack has to change, and none of it has to leave your hardware.
It's built by one person — Landon Horling — which explains both the pace and the shape of it. v1.0 is out with a free-forever developer key, and v1.x has been shipping something substantial most weeks since.
Software this young is usually marketed as if everything in it were finished. UPtrim labels itself instead. Features that ship in shadow mode run beside the live path, measure themselves against it, and only take over when the numbers earn it — the memory brain and the storage engine are both doing that right now, and the site says so rather than claiming they're live. Features that ship switched off are described as one switch away, not as default behaviour.
The settings reference is generated from the code's own shipped defaults, so it cannot drift from reality — an unclassified setting fails the build. There's an adversarial security review behind the hardening claims, with every confirmed finding fixed. And there is no analytics on this website, no telemetry in the product, and nothing phoning home about what you asked your AI. That isn't a privacy feature we're planning; it's just what's true.
one developer, one very long changelog
In order, for once. What shipped, what's shipping, and what's honestly still just a version number.
The spine: automatic fact extraction, hybrid keyword-plus-embedding retrieval, per-user isolation, 33+ file formats you can just ask about, multi-backend routing, and the admin dashboard. Free Developer key, no expiry, same key for everyone.
The platform is live and under active development. What's next lands when it's ready.
No date and deliberately no feature list. The interesting work is all inside v1.x right now: the memory brain graduating out of shadow mode on its own evidence, the agent system earning a default-on, and the multi-node fleet becoming something you can buy without sending an email first. v2.0 gets a page when it means something specific. Ideas and complaints are welcome on GitHub in the meantime.