The Problem
Say, we are using Gemini Deep Research to shoot a feasibility study task, 10 mins later, with a couple of follow-up discussions;
Having some long-term discussions with ChatGPT;
And using Claude Code and/or Cursor to build things;
We switch between them back and forth every day.
Everything works smoothly within each tool, and we're fortunate to take advantage of what today's GenAI models and products have to offer.
Until...
We realized it was a shame that some of our conversations were so insightful—moments we celebrated in the moment—yet later, we couldn't recall those brainstormed ideas, iterations, or discoveries.
Or, at the very least, it's that same helpless feeling you get when you have to explain complex backgrounds and context—again and again—not just to different people, but to your AI agents, or even across different threads with the same agent.
What If
What if we could collect all our most important memories—like the Pensieve in Harry Potter, where Dumbledore stores the key moments of his life, ready to revisit them at any time? And when he wants Harry to learn from those memories, Harry simply jumps in and experiences the context firsthand.
What if we had our own Pensieve—not just something we trust, but something effortlessly accessible by both humans and AI tools?
Nowledge Mem
We felt those pains, and were dreaming of the Pensieve-like thing that we can both trust and with ease to be accessed by human and AI tools, so we built:
Nowledge Mem
Nowledge Mem is a local first, graph augmented personal context manager.
Local First
Like Dumbledore won't hand over the Pensieve to death-eaters, our day-to-day thoughts and ideas should be persisted in a way that we can trust.
Plus, as of 2025, running AI workflows offline is not just because-i-can but some users actually need it.
So we persist, compute by default purely in the local device, and access the remote LLM only when user explicitly ask for it.
Graph Augmented
With memoris accumulated, we found it's crucial to be able to actually distill, extract and finally link them, so in a way that we can actually make sense of them, by solving the "needle in the haystack", "connecting the dots" and "getting the big picture" challenges.
Or, like how our brain works, we are using the Graph to connect different pieces of memories and Graph Algorithms to help us compute and recall the context when we need it.
How It Looks Like
In the initial version, Nowledge Mem is a native app on Desktop.
Where you can search or create memories within your AI agents like Claude Code, Cursor, ChaptWise, or any other tools speaks "MCP", the "USB" in the AI world. The Agent, say Cursor, can either search memories/create memories of recent findings, explicitly by users' request, or autonomously during the agent loop execution.
Not just do in the agentic way, if you would like to quickly consume memories from any app manually, Nowledge Mem comes with a Launcher, call it with ⌘ + ⇧ + K, you could pick any memories and paste to anywhere you are working on:
Specially, with nature of Graph in Nowledge Mem, we could explore and pick our memories in a more Graphy way, where you could see:
- first, we get a sub-graph of memories and entities, and the connections between them;
- then, we decided to explore some neigbors of the selected entities, thus to get more context expanded;
- finally, we could ckerrypick the memories and entities we want to use, and paste them to where we are working on.
What's Next
If you are interested in experiencing Nowledge Mem, it's still under invited alpha, you could join the waitlist via Nowledge Mem, read the documentation, or follow us on 𝕏 via @NowledgeMem and @NowledgeLabs.