The space between scattered notes and one brain.
NeuroBase Labs works on four hard problems on the path to a memory layer that lives in your Claude Code. Here is how it actually fits together — and what we're still figuring out.
How a meeting becomes an answer in your Claude Code.
Five stages turn a raw conversation into something you can question. Nothing is magic — just a pipeline that respects your data and shows its work.
Your tools
Connect the notetakers and calendar you already use. Lookout reads across all of them — no rip-and-replace, no new habit.
Ingest & extract
Every transcript is parsed into structured facts: who said what, what was decided, what was promised, and when.
Memory graph
Facts link into one private knowledge graph — decisions, commitments, people, dates — that survives the meeting and compounds over time.
MCP server
A secure, permissioned Model Context Protocol server exposes the graph. You own the brain; nothing leaves it without your say.
Your Claude Code
Recall and act from inside Claude Code — ask a question, get a grounded answer traced to the exact source moment. Works anywhere MCP is supported.
Fig 01 — the Lookout pipeline, source to recall
Four problems we're working on.
Click any area to open it.
An hour of conversation is dense, messy, and human. Turning it into a durable grapha machine can reason over — not just a wall of text to search — is the core problem.
Our approach has three moving parts:
- Extraction — an LLM pass turns each transcript into typed facts: decisions, commitments, people, dates, topics.
- Entity resolution — the same person, project, or commitment is recognised across many meetings, so the graph compounds instead of fragmenting.
- A traversable representation — facts become nodes and edges (who decided what, what depends on what), retrieved by graph walks and vector similarity, every node carrying its source.
Open question: how do we keep the graph accurate when later meetings contradict earlier ones? Memory has to forget and update, not just accumulate.
The valuable, difficult part isn't reading one notetaker — it's reading all of them into one coherent memory. Fireflies, Granola and Otter each have their own formats, speaker labels, and overlapping recordings of the same meeting.
- Normalise wildly different transcript formats into one schema.
- Reconcile speakers and de-duplicate overlapping records of the same call.
- A connector per source, so adding a tool never means re-architecting.
This is the wedge: incumbents structurally cannot aggregate a competitor's data.The seam between tools is exactly where the value lives — and where we build.
Open question: gracefully handling partial, low-quality, or auto-generated transcripts without polluting the graph.
Memory is only useful if a model can act on it. We expose the graph over the Model Context Protocol— so your Claude Code (and anything else that speaks MCP) can recall and reason, with you in control.
- A small set of MCP tools: search memory, fetch a decision, trace a commitment, draft from context.
- OAuth 2.1 and scoped permissions — the user owns the brain, the vendor does not.
- Not just read: safe, auditable actions the agent can take on your behalf.
Open question: where the line sits for autonomous write-back — what an agent may do unattended vs. only with a confirm.
An answer about your own history is worse than useless if you can't trust it. Every recall has to trace back to the exact source moment it came from.
- Per-claim citations — click through to the meeting and timestamp.
- Calibrated confidence, surfaced honestly.
- We would rather say “we don't know” than confidently invent your past.
Open question: how to measure groundedness well enough to improve it — a metric for “did this answer actually come from the record?”
We're a lab.
we publish what we learn.