EmbeddingGemma 2: multimodal semantic search and the future of agent memory

By Turing Post · Mon, 12 Oct 2026.

This issue of Turing Post focuses on EmbeddingGemma 2, the embedding model Google released last week, and ties it to the week's theme: agent memory (and, by extension, our own). The author starts from a personal premise. She forgets a lot of things, but she says the real problem is not that. It is the sheer amount of information we have to take in.

According to the newsletter, EmbeddingGemma 2 can find a video from a spoken description, search audio recordings with a written question, or locate images that look like a sketch. It has 740 million parameters, can run on-device and is released under the Apache 2.0 license. The author gives the example of a pile of badly named videos: you record a voice note asking for a specific moment, and a search app built on the model returns the right clip. She sees this as useful both for her agents' memory and for her own.

The piece explains what an embedding model is. It makes information searchable by meaning, turning content into numerical representations that place related things close together. What stands out about EmbeddingGemma 2 is that it does this across text, code, images, audio and video. A question and a video clip can therefore match even when the two files have little in common.

This addresses a specific problem. Many search systems first turn audio, images and video into text through transcripts and captions, then search those descriptions, so any details left out become harder to find. A transcript might capture a speaker's explanation of a chart but miss the relationships the chart itself shows. EmbeddingGemma 2 processes the media directly, giving access to visual and audio information that a description could lose.

The author acknowledges that multimodality is not new. Meta's ImageBind linked six modalities in 2023, including depth and thermal data, but it was released under a non-commercial license. What makes Google's offering interesting is the combination: a compact, commercially usable model with tools to run it on everyday devices. Google cites about 567 MB of active RAM for the full quantized model on a Pixel 11 Pro.

The engineering also lets developers choose how much of the model they need. Text and code use a 270-million-parameter component, vision adds 170 million and audio another 300 million. These configurations share a compatible embedding space. A written query can use the small text component to search media indexed with the full model, and audio search can be added later without recomputing existing text embeddings.

Another feature is the “Matryoshka” part. Like Russian nesting dolls, a smaller, still-useful representation sits inside the larger one. The model is trained so that the first 512, 256 or 128 numbers of its 768-dimensional embedding still work as a search representation after normalization, so developers can choose how much space to give each embedding. At 256 dimensions, the vectors take up a third of the storage, and Google says it keeps about 95% of retrieval quality across image, video and speech at that size. Going down to 128 saves more space but loses considerably more multimodal quality. Matryoshka Representation Learning dates back to 2022, so it is not unique to this Gemma. Still, the author finds this an interesting implementation, because a growing searchable history has to fit somewhere, and smaller vectors make it more practical to keep that index on the device.

Apps are already available to try. Google AI Edge Gallery includes media search and a tool for finding moments in videos. Google's Foresight app combines EmbeddingGemma 2 with Gemma 4 for retrieval and reasoning over local files: one model finds the material and the other works with it. For an agent, this could mean searching screenshots, code, recordings and earlier task reports together, and checking the original material before acting. A detail left out of a written summary yesterday might still be in a screenshot.

The closing reflection turns to human memory. The author compares a person's daily information intake in 1826 with today's. By her estimate, information then came mostly from face-to-face conversation and, for a highly literate person, a local newspaper, a few pages of a book or a letter, well under 2,000 to 5,000 written words a day. Today, by a very rough approximation, we take in at least 100,000 words a day. Our brains, she notes, are more or less the same as those of Homo sapiens some 300,000 years ago. That is why she sees the debate about agent memory as a deeper debate about our own memory. Superintelligence could be a way to hand off part of the work, such as searching and remembering where everything is, to a technology that never rests.

The newsletter also announces that on October 13 the author is chairing and curating O'Reilly's “Expert Showcase: Agent Memory.” It is a free online conference on robust, scalable memory architectures, focused on how agents keep useful context, retrieve what matters and learn from past interactions. Turing Post's lineup this week is dedicated to memory: on Wednesday, “State of Agent Memory” (AI Builds AI); on Friday, a guide to 3D Gaussian Splatting and how robots remember what is no longer in view; and on Sunday, the Library, with research on Looped Transformers.