AI Glossary

Term	Definition
Generative AI	AI systems that can create new content (text, images, audio, etc.) based on training data and prompts.
Machine Learning	A subset of AI focused on algorithms that improve through experience.
Deep Learning	A type of machine learning using neural networks with multiple layers.
Natural Language Processing (NLP)	AI field concerned with the interaction between computers and human language.
Computer Vision	AI field that trains computers to interpret and understand visual information.
Reinforcement Learning	Machine learning method based on rewarding desired behaviors and punishing undesired ones.
Neural Network	A computing system inspired by biological neural networks, forming the basis of many AI models.
AGI (Artificial General Intelligence)	AI systems capable of matching human intelligence in both cognitive tasks (reasoning, learning) and physical interactions (sensorimotor skills, environmental manipulation).
ASI (Artificial Super Intelligence)	AI systems that surpass human intelligence across all domains, including both cognitive abilities and physical world interactions, with capabilities beyond human comprehension.
SOTA (State-of-the-Art)	The best performing model or system in a specific domain or task. SOTA models set new records on recognized benchmarks.
BYOK	Bring Your Own Key
Embedding	A numerical representation of the meaning of a piece of content (text, an image, or audio) produced by an embedding model. Inputs with similar meaning produce embeddings that sit close together, which is what makes vector search possible.
Vector	A list of numbers that is the concrete form an embedding takes (for example, an array of 768 floating-point values for a 768-dimensional text embedding). "Vector" and "embedding" are often used interchangeably, but the vector is the data structure that stores the embedding.
Embedding Model	A model that converts raw content into a vector. Different content types typically use different embedding models (e.g., text versus image), and models can differ in the number of dimensions they output.
CLIP (Contrastive Language-Image Pre-training)	A multimodal embedding model with separate encoders for images and text that maps both into the same vector space, so a picture of a cat and the word "cat" land close together. This lets a system search images using natural-language text queries.
Whisper	A speech-to-text model that transcribes spoken audio (such as a voice note) into text locally, on-device. The resulting text is embedded and searched like any other text memory; the audio itself is typically not stored.
Distance Metric	The mathematical function used to measure how close two vectors are. Cosine distance (or cosine similarity) is a common choice, comparing the angle between two vectors rather than their magnitude.
Vector Search Engine	Software (such as Qdrant Edge) that indexes vectors into a searchable structure, often a navigable graph, so that given a query vector it can find the closest stored vectors extremely quickly, sometimes in under a millisecond.
ANN (Approximate Nearest Neighbor) Search	The technique vector search engines use to find vectors close to a query vector without comparing against every stored vector. It trades a small amount of accuracy for large gains in speed, which makes fast search over large memory stores possible.
Retrieval	Returning the vectors (and their associated content) closest to a query vector. Retrieval always returns the nearest match it can find, even a poor one, which is why a similarity threshold is needed to decide whether a result counts as a real match (see Recognition).
Similarity Score	A number produced by comparing a query vector to a stored vector using the chosen distance metric. Higher scores indicate the two vectors, and the content they represent, are more alike in meaning.
Query	A piece of text, an image, or a transcribed voice note that is embedded into a vector and used to search a memory store. Because it goes through the same embedding model as the stored memories, the search compares meaning rather than exact text.
Qdrant Edge	An open-source, on-device vector search engine used to build and query a local memory store, without requiring a separate database server.
Collection	A named set of points that share the same vector configuration (dimensions and distance metric) and can be searched together. Also called a "shard" when referring to the on-disk directory that holds that memory store.
Shard	The directory on the device where a collection's memory store physically lives. In an on-device setup, creating a shard and creating a collection describe the same underlying store.
Point	The basic unit of stored data in a vector search engine: a record made up of one or more vectors plus an optional payload of metadata, identified by a unique ID. Storing a memory means creating and uploading a point.
Payload	Structured metadata attached to a point, such as source type, category, price, location, or timestamp. It travels alongside the vector and can enrich results or filter a search separately from the similarity comparison.
Field Index	An index built on a specific payload field (such as category or price) that enables fast filtering on that field during search, so a query can be restricted to points matching certain metadata conditions.
source_type	A payload field recording where a memory originated (e.g., text, voice, or photo), so memories from different input channels can be distinguished and filtered even though they are searched together.
Similarity Threshold	A cutoff score above which a retrieved match is treated as confirmed rather than merely "closest available." Retrieval always returns a nearest neighbor; the threshold decides whether that neighbor is close enough to trust.
Recognition	Deciding that a newly observed object or memory is the same as something previously taught, based on its similarity score clearing the threshold. Recognition is retrieval plus a threshold check.
False Positive	A reported match (similarity above threshold) between two things that are not actually the same, indicating the threshold may be set too low.
False Negative	A missed match (similarity below threshold) between two things that actually are the same, indicating the threshold may be set too high.
Threshold Calibration	Choosing a similarity threshold by testing it against a labeled set of known matches and non-matches, then picking a cutoff, often the midpoint between the lowest true-match score and the highest false-match score, that balances false positives against false negatives for the use case.
Capture	The first stage of the memory lifecycle: an experience (a camera frame, a spoken note, or typed text) is observed and prepared for storage by being turned into an embedding.
Recall	Searching a memory store with a question or image to retrieve related past memories. The "read" side of the memory lifecycle, built on retrieval and, where relevant, recognition.
Forgetting	Permanently removing a memory (a point) from the memory store, typically because the underlying object or fact is no longer relevant or was mislabeled. Distinct from soft forgetting, which lowers a memory's ranking without deleting it.
Soft Forgetting	Reducing the influence of older memories on search results over time, without deleting them, by combining a freshness weight with the similarity score so more recent memories tend to rank higher.
Half-Life	A configurable time span in soft forgetting that sets how quickly a memory's freshness contribution decays. With a seven-day half-life, a memory's recency boost roughly halves every seven days.
Freshness Weight	A parameter controlling how much recency influences a memory's final ranking relative to its raw similarity score, letting newer memories outrank older, otherwise-similar ones.
Metadata Filtering	Restricting a vector search to points whose payload matches specified conditions (such as category or price range), applied alongside, not instead of, similarity ranking.
Multimodal Memory	A memory store that combines more than one content type, such as text, images, and transcribed voice, in a single searchable store, typically by giving each type its own vector type and embedding model within the same collection.
On-Device Processing	Running the embedding models, vector search, and storage locally on the device rather than on a remote server. Enables offline use, generally faster and more reliable recall than a network round trip, and keeps private memories on the device.
Cloud Sync	The optional ability of an on-device memory system to synchronize its local store with a cloud server when more compute or storage is needed or memories must be shared, while still supporting fully local use.
Headless Mode	A device configuration in which the system runs without any monitor, keyboard, or mouse attached, relying on a separate interface such as a phone connected over direct Wi-Fi.
Latency	The time it takes for a search query to return results. Vector search engines are designed so latency grows much more slowly than the number of stored memories, which lets them scale to very large memory stores.