Reachy Remembers: Giving a Robot a Local Semantic Memory with Qdrant Edge

Reachy Remembers equips the Raspberry Pi–based Reachy Mini robot with a private, on-device semantic memory using Qdrant Edge, allowing it to recognize returning people, recall conversations beyond a language model’s context window, and describe scenes it observed earlier without sending stored data to the internet. The system combines Whisper speech transcription, camera-based object and face detection, Gemma 4 dialogue handling, SigLIP2 image embeddings, and bge-small text embeddings, while storing conversational exchanges, visual frames, faces, and fixed robot knowledge in separate local vector-search shards. Testing revealed that effective memory required careful tool descriptions to distinguish past-focused recall from requests to inspect the current scene, as well as score thresholds and an “empty frame” comparison to prevent vector search from returning irrelevant nearest matches for vague questions. When no specific memory matches, the robot selects diverse stored images to provide a broader account of its day, while face recognition uses multivectors and confidence thresholds to greet known people or ask unfamiliar visitors for their names. Although the demonstration runs large AI models on a nearby laptop due to memory limitations, all memory storage and searches remain on the robot, with searches taking roughly half a millisecond across 500 memories. The project’s broader lessons emphasize thresholding retrieval, treating tool prompts as part of the retrieval system, and organizing vector stores around access patterns, with plans to release the capability as a Reachy Mini app.

Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.