Reachy Mini robot gets a local semantic memory with Qdrant Edge

Why a robot needs its own memory

Reachy Mini runs on a Raspberry Pi with 4 GB of RAM. Qdrant's authors, Sasha Denisov and Chadha Sridi, wanted it to behave like a good host: greet you by name the second time you visit, remember what you talked about and report what it saw while you were away.

A robot driven by a language model cannot do this alone, they write. The model only knows what is in its context. Restart the robot and it knows nobody, talk long enough and the start of the conversation drops out, and nothing in it can be searched.

The memory is stored locally because it consists of pictures of a room and the people in it. According to Qdrant, that data should not leave the house unless the owner decides so. Local storage also means a search takes a fraction of a millisecond and still works when the Wi-Fi does not.

How the loop and the memory work

Everything Reachy does runs through one loop. Whisper turns speech into text, a detector and a face model read the camera, and Gemma 4, the dialog model, either answers or calls one of three tools: remember to search memory, camera to take a fresh picture, and move to make a gesture.

Memory is written without a tool call: an exchange when the context fills up, a frame when the objects in view change or Reachy is asked to look, and a face when it learns a name. SigLIP2 embeds pictures and bge-small embeds text.

Qdrant Edge is the Qdrant vector search engine shipped as a library loaded into the application's own process, with no server and no network hop. Data is kept in shards, which are self-contained directories on disk. Reachy uses three:

• memory/ holds exchanges and frames, separated by a kind field indexed as the tenant key. A frame carries two named vectors, one from SigLIP2 and one from bge-small, so it can be found by picture or by words.
• people/ holds one point per person, with every shot of their face in a multivector. MaxSim scores the camera face against each person's closest shot. Above 0.35 Reachy greets by name, below 0.25 it asks who you are, and in between it waits.
• knowledge/ holds 62 facts about Qdrant and the robot, each stored with the questions it answers. The robot never writes to it. It is restored at every start from a 0.7 MB snapshot, the same format a Qdrant Cloud collection produces.

When Gemma calls remember, it sends a query and an about label. The query is embedded first, and the label narrows which shards are searched. Each shard has its own score gate so a weak nearest neighbor does not count as an answer.

The "What did you see today?" problem

The team says one question taught them most. Asked a few turns into a conversation, Reachy turned its head, took a new picture and described the room, instead of searching memory. A small model with tools picks one from wording alone, and "what did you see" sounded like "look."

The fix was in the description of the remember tool, which now states that anything in the past tense belongs there and quotes example questions. According to Qdrant, this raised correct tool choice from 23 to 26 out of 26 test questions without costing the camera tool any of its own.

A second failure followed. "What did you see?" names nothing, yet a vector search always returns a nearest neighbor. Frames are searched through a short stored sentence such as "I saw person", so the question matched them on the verb alone and scored 0.704, above the 0.66 gate. The fix was to add an empty frame, "I saw something", that every result must also beat. Per the authors, every test question that named nothing scored under it, and every question about something present scored over it.

When nothing beats the empty frame, Reachy samples its day instead. It starts from the stored frame whose image vector is nearest the question, then repeatedly adds the frame least like those already chosen. The newest four frames of one evening were the same picture four times; picked this way, they are four different scenes.

What stays on the robot

The authors are precise on one point. Gemma 4 alone weighs 2.5 GB before its KV cache, and the robot has about 3 GB of RAM free, so in this demo the models run on a laptop beside Reachy over the local network. Storage and search stay on the robot: every vector and payload lives on its disk, and every search runs there, in about half a millisecond for the 500 memories of a demo run. Recalled pictures and text do reach the laptop, where Gemma reads them. Nothing goes to the internet.

The conversation memory also outlives the model's context window. The demo caps context at 1,000 tokens, and once it is full the oldest half moves into memory instead of being dropped, so "what did we talk about?" still gets an answer an hour later.

Lessons and next steps

Qdrant says the lessons apply to any device that keeps its own memory, such as a phone or a pair of glasses. A search always returns something, so each store needs its own score gate tuned on real questions, plus an empty record to beat. The tool description is part of the retrieval system: a sentence about tense did more for recall than any change to the index. Shards should be split by how data is written and searched, not by what it is.

Next, the team is packaging Reachy Remembers as a Reachy Mini App so owners of the robot can add on-device memory in one install. Until then the code is in the reachy-edge-memory repository, and a technical writeup covers the full model list, the shard configuration and how search scales to 250,000 memories on the same Raspberry Pi. Qdrant says Qdrant Edge is free to use.