Kit For AI
The memory layer for AI agents
83 followers
The memory layer for AI agents
83 followers
Persistent memory and grounded knowledge for AI agents — native MCP tools your agent calls directly. Drop in any file or URL; skip building and babysitting a RAG stack. One API, works with any model. Free to start.
Kit For AI
Born from my own daily pain: my agent forgot everything I told it yesterday, half my knowledge lived in formats it couldn't read (PDFs, YouTube, docs sites), and the "fix" was babysitting a 5-service RAG stack. So I built it once, properly: any file/URL in clean, searchable knowledge out, plus remember/recall as native MCP tools. One install, any model.
Same problem everyone has "I wish my AI knew my stuff" solved as a drop-in layer.
honestly this looks super useful, been wanting something like this for ages. one thing though, it would be really helpful if there was a way to set automatic refresh intervals for ingested URLs so the knowledge stays current without having to manually re-upload stuff all the time
Kit For AI
Thanks @furkanbeki61828 for the feedback! 🙌 We already support automatic refresh intervals for ingested URLs ✅. We're also adding an interactive mode for URLs in the next release, so stay tuned!
vibecoder.date
How do you mitigate the inherent limitations of MCP when it comes to context window usage especially on large scale projects where even a million token context window seems minuscule?
Kit For AI
Really good question 🙌
The shift that made this click for us: the context window isn't storage, it's a desk. You don't put the whole filing cabinet on your desk — you grab the two pages you need right now.
So on a big project, the agent never loads the codebase or the docs. It calls a tool, and gets back only the top few relevant chunks — reranked and capped to a fixed token budget. A 500-doc project and a 5-doc one send roughly the same amount to the model. The library grows, what hits the context doesn't.
Three things keep that slice sharp:
→ hybrid search (meaning + exact terms) so the right chunk wins, not just a nearby one
→ a hard token budget before anything reaches the prompt
→ memory that retires old facts when they change, so it doesn't slowly rot
Honestly the hard part isn't window size, it's retrieval quality. A clean 4k-token slice beats a stuffed 1M window on accuracy and cost.
Where it still gets tough is reasoning that spans the whole project at once that's what we're chasing next.
What are you building? Curious if it's that case 🙂
vibecoder.date
@l33tcy I've been thinking about this problem from a tangential perspective.
I am tempted to build my own agent harness. Thing is semantic/vector search lacks directionality and representation of the 'how' of relevancy.
Term search is good. What really has me curious is making a token efficient mechanism to manage context by compressing and decompressing intelligently as you build.
Some things are obvious, like replacing a code change or diff with a line summary and file reference, but mapping out architectural concepts and viewing things at the right level of detail is... it's hard.
The remember/recall-as-native-MCP-tools approach is what makes this feel usable versus babysitting a 5-service RAG stack. Two setup questions: is the knowledge store hosted by you, or can it run local/self-hosted so the data stays on my machine, and is memory scoped per-project so pointing it at two clients' docs can't bleed one into the other on recall? Also, when I drop in a URL, is that a one-time snapshot or does it re-crawl when the page changes?
Real talk 🙌 this is exactly what I needed. I've been struggling for a while with AI models forgetting context every single session, having to re-explain the same things over and over got exhausting . Kit For AI finally fixed that for me: I drop in a file or URL once, and my agent actually remembers it going forward, no extra setup, no complicated RAG pipeline. Huge thanks to the team for building something this simple and practical 🙏 it genuinely saves me hours every week.
Quick question though: how does the near-duplicate memory dedup work when the same fact gets updated over time, does it version the old memory or just overwrite it ?
Congrats on the launch. The memory-as-MCP-tools framing is practical because it keeps the agent from treating the context window like storage. One thing I’d want as a builder is provenance on recall: when a memory is used, can the agent surface which file, URL, or chunk it came from and whether that source has changed since ingestion?
One thing I'd love to see is a way to scope which agent or project owns a particular memory or knowledge doc, so different agents using the same Kit instance don't accidentally overwrite or pull each other's context.