wherebase
Stealth · private beta with early partners

Spatial memory, with time attached

Memory for things that move.

Robots, wearables and agents forget a place the moment they leave it. Wherebase keeps where and when together, so a machine can remember what was here, notice what changed, and know what usually comes next.

What we're building

Three questions a map can't answer

What was here, when, and is that normal?

01 · Where

A memory with coordinates.

Every observation is anchored to a point in space, at the scale of a desk, a warehouse or a city. Ask about a place, not a file.

02 · When

Time as an axis, not a column.

What was on this shelf yesterday. What's usually by the door on Tuesdays. Which route this machine has taken before, and where it went next.

03 · What changed

Remember the surprises.

The system predicts what it expects to find before it looks. Only the difference is worth keeping, so memory stays small and attention goes where it matters.

Roadmap · updated November 2026

What's built, what's next.

We build in the open where we can and keep quiet about the rest. This is the honest state of it: things you can hold, things on the bench, and what comes after. Numbers are from our logs, not a deck.

Working today

  1. The memory engine

    Four axes, x, y, z and t, indexed together on a space-filling curve at several resolutions, so "what was near here around then" is one lookup. Storage is bounded and ages like memory: raw moments consolidate into summaries instead of piling up. Before it looks, it predicts what it expects to find, and the difference is a calibrated "this is new" signal.

    DemoA simulated warehouse robot patrols, builds memory, then answers place-and-time and "have I seen this?" questions. Five steps, runs in-process, no keys.

  2. Cloud, at the edge

    A multi-tenant service on Cloudflare's network: projects, keys, spaces, device tokens, one isolated memory per space, and spoken answers generated at the edge. The same client code runs against a local store or the cloud.

    DemoCloud mode: a board on the neck talks to the service with a device token and no laptop in the loop.

  3. A pendant you can wear

    An ESP32-P4 board with a camera, microphone and speaker streams frames at 2 fps over one keep-alive connection. Nothing heavy runs on the board. Lose the network and it keeps going: frames queue on the microSD card and upload later with their original timestamps, and a small on-device model still understands "where is my cup".

    DemoPush the button, ask where your cup is; the answer plays from the board about 1.3 s after you let go. Outage test: six hundred queued frames drained while live frames kept flowing.

  4. Voice, end to end

    Speech in, a spatial answer out, in words: "on the kitchen table". Speech recognition and synthesis run on the host or at the edge, so the same question works from a pendant, a phone or a headset.

    DemoRoom assistant: a phone's camera watches a room, you ask out loud where things are, and it tells you.

  5. More devices, one protocol

    The same host protocol runs on ESP32-S3 camera boards and on a Raspberry Pi 5 with a stereo depth camera, where objects get metric positions and visual tags anchor rooms.

    DemoHandheld rig: depth gives distances, so answers are "two metres ahead, on the left", not just "somewhere in the kitchen".

  6. Memory for agents

    An MCP server lets LLM agents store and recall spatial memories as tools, and adapters feed world models (V-JEPA 2, DreamerV3) from the same store.

On the bench

  1. Seeing on the device itself

    Moving detection from the host onto the pendant so it can work with a slow or absent link. Measured so far: decoding a frame is cheap, the on-chip detector is not yet, and we are testing dedicated vision silicon against it.

  2. Your objects, not a fixed list

    Open-vocabulary detection so "where are my keys", "my glasses" and "the pill box" work, not just the eighty categories stock detectors know. Every board can also record frames to fine-tune a detector on your things.

  3. The companion app

    Voice-first setup so a blind person can claim and configure a device without a laptop, and a carer's view with rooms, a timeline and whether the device is alive. Designed; being built.

  4. One memory, many devices

    Several devices writing into the same space, so a fleet of robots or a household of wearables shares what any one of them has seen.

Next

  1. A pendant that stands alone

    Cellular link, wake word, battery life measured rather than guessed, and an enclosure. The current board is a development kit with a camera ribbon.

  2. Robots, properly

    A ROS 2 node that turns odometry and camera embeddings into memories, a LeRobot loader, and Isaac Sim and Habitat examples with real embeddings.

  3. Private beta

    With blind and visually impaired testers first, because they will tell us fastest when the memory is wrong.

  4. Guarantees, not vibes

    False-alarm bounds on the "this is new" signal and a public benchmark for memory in world models.

Maps say where things are. Logs say what happened. Nothing says what was where, when, and whether that's normal. We're building that.

One memory, shared by every device in a fleet and every agent that reasons about it. Small enough to run on a pendant, fast enough for a robot, honest enough to say "I haven't seen this before." Built first with people who navigate the world without sight, because they notice when a machine's memory is wrong.