The MemX Glossary
Clear, sourced definitions for the ideas behind AI memory, search and knowledge management. The concepts a second brain is built on, explained for humans and the AI tools that cite them.
AI Foundations
Large Language Model (LLM)A large language model (LLM) is a neural network, typically a transformer trained on massive text corpora, that learns the statistical patterns of language by predicting the next token, and uses that learned distribution to generate and process human-like text.Transformer (Neural Network Architecture)A transformer is a neural network architecture, introduced in the 2017 paper "Attention Is All You Need," that processes sequences using self-attention instead of recurrence. It underpins most modern large language models, vision transformers, and multimodal systems.Natural Language Processing (NLP)Natural language processing (NLP) is the subfield of artificial intelligence and computer science concerned with enabling computers to read, interpret, and generate human language, spanning tasks such as tokenization, parsing, classification, named entity recognition, translation, and summarization.Generative AIGenerative AI is a class of machine learning systems that learn the underlying distribution of training data and sample from it to produce new content, including text, images, audio, video, and code, rather than only classifying or scoring existing inputs.Language ModelingLanguage modeling is the task of learning a probability distribution over sequences of tokens, typically by predicting the next token given the preceding context. This single objective is the statistical foundation on which every large language model is built.Convolutional Neural Network (CNN)A convolutional neural network (CNN) is a deep learning architecture that processes grid-structured data, such as images, using learnable filters that slide across the input to detect spatial patterns. Stacked convolution and pooling layers build features from edges up to objects.Activation FunctionAn activation function is a nonlinear transformation applied to a neuron's weighted-sum output, enabling neural networks to learn complex, nonlinear relationships. Without it, stacked layers collapse into a single linear map. Common choices include ReLU, sigmoid, softmax, and GELU.Tokens (in Large Language Models)A token is the atomic unit of text that a large language model reads and generates: a chunk that may be a whole word, a subword fragment, a single character, or punctuation. In English, roughly 100 tokens equal about 75 words.Tokenization (and Tokenizers)Tokenization is the process of splitting raw text into discrete units (tokens) that a language model can map to integer IDs. A tokenizer is the tool that learns and applies these splits, usually as subword pieces produced by algorithms like BPE, WordPiece, or Unigram.Multimodal AIMultimodal AI refers to systems that process and relate more than one type of data, such as text, images, audio, and video, by mapping each modality into a shared representation so information from one can inform reasoning or generation in another.Reinforcement LearningReinforcement learning (RL) is a machine learning paradigm in which an agent learns to make decisions by interacting with an environment, taking actions and adjusting its behavior to maximize cumulative reward over time.Mixture of Experts (MoE)Mixture of Experts (MoE) is a neural network architecture that replaces a dense layer with many parallel sub-networks called experts, plus a router that activates only a small subset per token. This lets a model hold huge total parameters while keeping per-token compute low.Machine Learning (ML)Machine learning (ML) is a branch of artificial intelligence in which a program learns to perform a task from data and experience rather than from rules a human writes by hand. The system improves its performance as it sees more examples, discovering patterns it was never explicitly programmed to follow.Deep LearningDeep learning is a subfield of machine learning that uses neural networks with many stacked layers to learn a hierarchy of features directly from raw data. The word "deep" refers to the number of layers: each layer transforms its input into a slightly more abstract representation, and stacking many of them lets the model learn complex patterns without hand-engineered features.Neural NetworkA neural network is a machine learning model made of layers of interconnected units called neurons, each computing a weighted sum of its inputs followed by a nonlinear activation. It learns by adjusting those weights to reduce prediction error on training data.Supervised LearningSupervised learning is a machine learning approach that trains a model on labeled examples, each pairing an input with its correct output, so the model learns a function that maps new inputs to predictions. Its two main tasks are classification (predicting a discrete category) and regression (predicting a continuous number).OverfittingOverfitting is when a machine learning model learns the training data too closely, including its noise and quirks, so it scores well on data it has seen but generalizes poorly to new, unseen data. You recognize it as a widening gap between low training error and high validation error.BackpropagationBackpropagation is the algorithm that trains neural networks by computing the gradient of the loss with respect to every weight. It runs a forward pass to produce a prediction and loss, then propagates the loss gradient backward through the layers using the chain rule, reusing each layer's stored values so the full gradient is computed in a single backward sweep.Gradient DescentGradient descent is an optimization algorithm that minimizes a loss function by repeatedly stepping the model's parameters in the direction opposite the gradient. Each update is parameter = parameter minus learning_rate times gradient. It is the core method used to train neural networks.Loss FunctionA loss function is a scalar measure of how wrong a model's prediction is compared to the true target. Training minimizes it: the loss produces the error signal that gradient descent and backpropagation use to update weights. Common choices are mean squared error for regression and cross-entropy for classification.Attention Mechanism (Self-Attention)The attention mechanism is a neural network operation that lets each token in a sequence weigh and pull in information from every other token. Self-attention computes a similarity score between each token's query and every token's key, normalizes those scores into weights, and uses them to build a weighted sum of value vectors, so each token's representation reflects the context most relevant to it.QuantizationQuantization is the process of lowering the numerical precision of a model's weights and activations, for example from 16-bit floats to 8-bit or 4-bit integers, so the model uses far less memory and runs faster during inference with little loss in quality.Diffusion ModelA diffusion model is a generative model that learns to create data by reversing a gradual noising process. It is trained to remove noise step by step, turning pure random noise into a coherent sample such as an image.Text-to-Image GenerationText-to-image generation is the task of synthesizing a new image that matches a natural-language prompt, where a generative model (today usually a diffusion model) is conditioned on a text embedding produced by an encoder such as CLIP or T5.Generative Adversarial Network (GAN)A Generative Adversarial Network (GAN) is a generative model in which two neural networks, a generator and a discriminator, train against each other in a minimax game: the generator learns to produce realistic synthetic data while the discriminator learns to tell real data from generated data. Introduced by Ian Goodfellow and colleagues in 2014, GANs powered a wave of photorealistic image synthesis before diffusion models overtook them for high-fidelity generation around 2021.Recurrent Neural Network (RNN)A recurrent neural network (RNN) is a neural network that processes a sequence one element at a time while maintaining a hidden state that carries information from earlier steps forward. The recurrence lets the same weights reuse past context to inform the current output, which makes RNNs suited to text, speech, and other ordered data.Named Entity Recognition (NER)Named Entity Recognition (NER) is the natural language processing task of locating spans of text that name real-world entities and classifying each into predefined types such as person, organization, location, date, or money.Unsupervised LearningUnsupervised learning is a branch of machine learning that finds structure in unlabeled data without any target labels, using tasks such as clustering, dimensionality reduction, and density estimation to discover patterns the data already contains.Ensemble MethodsEnsemble methods combine the predictions of multiple models so the group is more accurate than any single member, mainly by reducing variance (bagging), reducing bias (boosting), or learning how to blend models (stacking).Text-to-Video GenerationText-to-video generation synthesizes a coherent video clip directly from a natural-language prompt. It extends text-to-image diffusion to the time dimension, which makes keeping motion smooth and objects consistent across frames the central difficulty.Autoencoders and VAEsAn autoencoder is a neural network that compresses input into a low-dimensional bottleneck and reconstructs it, learning features without labels. A Variational Autoencoder (VAE) makes that bottleneck a probability distribution, turning the model into a true generator.Decision Trees and Random ForestsA decision tree is a model that predicts by recursively splitting data on feature thresholds; a random forest averages many decorrelated trees, each trained on a bootstrap sample with a random subset of features, to cut the high variance that makes single trees overfit.DropoutDropout is a regularization technique that randomly sets a fraction of a neural network's units to zero during each training step, which discourages units from co-adapting and reduces overfitting. At test time dropout is turned off and the trained weights are used directly.GPT and Decoder-Only ModelsA decoder-only model is a Transformer that uses causal (masked) self-attention to predict the next token from left-to-right context alone. GPT is the canonical example of this autoregressive architecture, which became the dominant design for generative language models.Text Classification and Sentiment AnalysisText classification assigns predefined categories to a piece of text, such as topic, spam, or intent labels. Sentiment analysis is its canonical case, predicting the polarity (positive, negative, or neutral) of an opinion expressed in text.Transformer Architecture (Deep Dive)A component-level walkthrough of how a Transformer block is assembled: positional encoding, the position-wise feed-forward sublayer, residual connections, layer normalization, how multi-head attention slots into the block, and how these stack into encoder and decoder layers.Memory PoisoningMemory poisoning is an attack in which malicious content is written into an AI agent's persistent memory so that it influences the agent's behavior in future sessions. It is the memory analog of prompt injection: instead of corrupting a single response, the attacker plants instructions or false facts that the model treats as trusted context every time it loads that memory.Temporary ChatTemporary Chat is a per-conversation mode in AI assistants like ChatGPT, Claude, and Gemini in which a session is not saved to your chat history, does not read or update your saved memory, and is not used to train the model. It reduces what the assistant remembers about a session, but it is not end-to-end encrypted, and providers still retain the messages for a limited window for safety and abuse monitoring.Rotary Position Embedding (RoPE)Rotary Position Embedding (RoPE) encodes a token's position by rotating its query and key vectors by an angle proportional to its position, so the attention dot product depends only on the relative distance between tokens. It is the default position encoding in modern LLMs such as LLaMA, Mistral, Qwen, and DeepSeek.FlashAttentionFlashAttention is an IO-aware algorithm that computes exact attention without ever writing the full attention matrix to GPU high-bandwidth memory. By tiling the computation into blocks that fit in fast on-chip SRAM and using an online softmax, it reduces memory traffic and turns attention's memory cost from quadratic to linear in sequence length.Grouped-Query Attention (GQA) and Multi-Query Attention (MQA)Multi-Query Attention (MQA) shares a single key-value head across all query heads, and Grouped-Query Attention (GQA) shares one key-value head across each small group of query heads. Both shrink the key-value cache that dominates LLM inference memory, with GQA preserving most of the quality of full multi-head attention.Multi-Head Latent Attention (MLA)Multi-Head Latent Attention (MLA) is an attention mechanism introduced in DeepSeek-V2 that compresses keys and values into a single low-rank latent vector per token. Only that small latent and a decoupled rotary key are cached, which DeepSeek reports cuts the KV cache by 93.3 percent versus standard multi-head attention.RMSNorm (Root Mean Square Normalization)RMSNorm (Root Mean Square Normalization) is a normalization layer that rescales a vector by its root mean square and applies a learned gain, without subtracting the mean. By dropping LayerNorm's re-centering step, it is simpler and faster while matching LayerNorm quality, which is why it is standard in modern LLMs.Feed-Forward Network / MLP Block (and SwiGLU)The feed-forward network (FFN), also called the MLP block, is the position-wise sublayer inside every transformer layer that applies two linear projections with a nonlinearity in between. SwiGLU is a gated variant that replaces the simple nonlinearity with a Swish-gated linear unit and is now standard in models like LLaMA and PaLM.Decoding Strategies (Greedy, Beam Search, Top-k, Top-p)Decoding strategies are the algorithms that turn a language model's predicted probability distribution over the next token into actual generated text. The main families are deterministic search (greedy, beam search) and stochastic sampling (top-k, top-p/nucleus), which trade off accuracy, diversity, and the risk of repetitive or degenerate output.Softmax FunctionThe softmax function converts a vector of real-valued scores (logits) into a probability distribution, where each output is between 0 and 1 and all outputs sum to 1. It is used in transformer attention to weight tokens and at the output layer to turn logits into next-token probabilities.State Space Models (SSM) and MambaState space models (SSMs) are sequence models that maintain a compressed hidden state and update it recurrently, scaling linearly with sequence length. Mamba is a selective SSM (the S6 layer) that lets its parameters depend on the input, matching transformer quality on language at the model sizes tested while avoiding attention's quadratic cost.Adam OptimizerAdam (Adaptive Moment Estimation) is a gradient-based optimization algorithm that adapts the learning rate for each parameter using running estimates of the first and second moments of the gradients. Introduced by Kingma and Ba in 2014, it is one of the most widely used optimizers for training neural networks.Vanishing and Exploding GradientsVanishing and exploding gradients are two failure modes in training deep neural networks where gradients shrink toward zero or grow uncontrollably as they propagate backward through layers. Both stall or destabilize learning, especially in very deep or recurrent networks.Vision Transformer (ViT)A Vision Transformer (ViT) is an image model that splits an image into fixed-size patches, treats each patch as a token, and processes the patch sequence with a standard Transformer encoder instead of convolutions.Vision Language Model (VLM)A Vision Language Model (VLM) is a multimodal model that combines a vision encoder, a projection module, and a large language model so it can take images and text together as input and produce text that reasons about both.Automatic Speech Recognition (ASR)Automatic Speech Recognition (ASR) is the technology that converts spoken audio into written text. Modern ASR systems use neural networks trained on large speech datasets, and their accuracy is commonly measured by Word Error Rate (WER).Cross-AttentionCross-attention is an attention mechanism in which the queries come from one sequence while the keys and values come from a different sequence, letting a model condition one set of representations on another, such as a decoder attending to an encoder or text attending to an image.Object Detection (YOLO / Bounding Boxes)Object detection is a computer vision task that finds and classifies every object in an image, drawing a labeled bounding box with a confidence score around each one. YOLO performs this as a single regression pass for real-time speed.Regularization (L1/L2)Regularization adds a penalty on model parameter size to the training objective so the model fits the data without growing weights excessively. L1 (Lasso) drives some weights to exactly zero for sparsity; L2 (Ridge) shrinks all weights smoothly toward zero.Gradient Boosting (incl. XGBoost)Gradient boosting builds a predictive model as an additive ensemble of weak learners, usually shallow decision trees, where each new tree is fit to the negative gradient (the residual errors) of the loss from the current ensemble. XGBoost is a regularized, second-order implementation.Principal Component Analysis (PCA)Principal component analysis is an unsupervised dimensionality reduction technique that projects data onto a new set of orthogonal axes, the principal components, ordered so that each captures the maximum remaining variance. Keeping the top components compresses data while preserving most of its structure.Logistic RegressionLogistic regression is a linear classification algorithm that models the probability of a class by passing a weighted sum of input features through the sigmoid function, then trains the weights by minimizing log loss.Latent Diffusion Models (LDM)Latent diffusion models run the diffusion process in the compressed latent space of a pretrained autoencoder instead of raw pixels. This cuts compute and memory dramatically while keeping image quality, and underpins Stable Diffusion.Classifier-Free Guidance (CFG)Classifier-free guidance steers a diffusion model toward its prompt by combining conditional and unconditional noise predictions at each denoising step. A guidance scale controls the tradeoff between prompt adherence and sample diversity.Word2Vec (CBOW & Skip-gram)Word2Vec learns dense vector representations of words from raw text using a shallow neural network. Its two architectures, CBOW and skip-gram, predict words from context or context from words, placing similar words near each other in vector space.
Training & Alignment
Fine-TuningFine-tuning is the process of taking a pretrained model and continuing to train it on a smaller, task-specific or domain-specific dataset, updating some or all of its weights so it specializes in a target behavior, format, or domain.Instruction TuningInstruction tuning is a supervised fine-tuning method that trains a pretrained language model on many examples of natural-language instructions paired with desired responses, teaching it to follow instructions and generalize to tasks it was never explicitly trained on.Reinforcement Learning from Human Feedback (RLHF)RLHF (Reinforcement Learning from Human Feedback) is a training technique that aligns a language model with human preferences by learning a reward model from human-ranked outputs, then optimizing the model's policy to maximize that reward while staying close to its starting behavior.LoRA (Low-Rank Adaptation)LoRA (Low-Rank Adaptation) is a parameter-efficient fine-tuning method that freezes a model's pretrained weights and trains small low-rank matrices injected into its layers, so a large model can be specialized while updating a tiny fraction of its parameters. The original 2021 paper reports cutting trainable parameters by up to 10,000x and GPU memory by 3x versus full GPT-3 175B fine-tuning, with no added inference latency.Jailbreaking (LLMs)Jailbreaking is the practice of crafting prompts that bypass a large language model's safety alignment to elicit content the model was trained to refuse. It attacks the model's learned safety policy, not the application around it.Transfer LearningTransfer learning reuses knowledge a model gained on a source task or domain to improve learning on a related target task, usually by pretraining on large data and then adapting the pretrained weights to a smaller target dataset.AI SafetyAI safety is the interdisciplinary field focused on preventing accidents, misuse, and other harmful consequences from AI systems, spanning concrete present-day risks and debated longer-term concerns about loss of control over highly capable models.AI AlignmentAI alignment is the problem of building AI systems that reliably pursue their designers' intended goals and values rather than some unintended proxy. It splits into outer alignment (specifying the right objective) and inner alignment (the trained model actually optimizing that objective).Direct Preference Optimization (DPO)Direct Preference Optimization (DPO) is a method that aligns a language model to human preference data by training directly on chosen-versus-rejected response pairs, replacing the separate reward model and reinforcement learning loop of classic RLHF with a single classification-style loss.Parameter-Efficient Fine-Tuning (PEFT)Parameter-Efficient Fine-Tuning (PEFT) is a family of methods that adapt a large pretrained model to a downstream task by training only a small set of new or selected parameters while keeping most pretrained weights frozen, cutting memory and storage costs versus full fine-tuning.Knowledge DistillationKnowledge distillation trains a small student model to mimic a larger teacher model, transferring the teacher's behavior through its full probability distribution so the student keeps much of the quality at a fraction of the cost.AI Bias and FairnessAI bias is the tendency of machine learning systems to produce systematically unfair outcomes for certain groups, while fairness is the study of formal criteria, audits, and mitigations used to measure and reduce that harm.Federated LearningFederated learning trains a shared model across many decentralized devices or organizations that keep their data local, sending only model updates to a central server that averages them. The raw data never leaves the client.Hyperparameter TuningHyperparameter tuning is the search for the configuration values set before training, such as learning rate, batch size, or regularization strength, that a model cannot learn on its own. The goal is to find settings that minimize validation error rather than fit the training data.Synthetic DataSynthetic data is artificially generated data, produced by simulation, generative models, or another model's outputs, used to train or evaluate machine learning systems in place of or alongside real-world data.Mixed Precision TrainingMixed precision training is a technique that performs most neural network computations in 16-bit floating point (FP16 or BF16) while keeping a high-precision copy of weights, cutting memory use and speeding up training on modern hardware without sacrificing accuracy.Constitutional AI (CAI)Constitutional AI (CAI) is a method developed by Anthropic for training AI models to be helpful and harmless using a written set of principles, called a constitution, instead of relying mainly on human labels for harmfulness. The model critiques and revises its own outputs against those principles.Reward Model (RLHF)A reward model is a neural network trained to predict how good a language model's output is, producing a scalar score that reflects human preferences. It is the component of RLHF that converts pairwise human comparisons into a learned reward signal used to fine-tune the main model.Proximal Policy Optimization (PPO)Proximal Policy Optimization (PPO) is a policy-gradient reinforcement learning algorithm that improves a model using a clipped objective, keeping each update close to the previous policy for stable training. It became the default optimizer for RLHF fine-tuning of large language models.Group Relative Policy Optimization (GRPO)Group Relative Policy Optimization (GRPO) is a reinforcement learning algorithm that removes PPO's separate value model. It estimates each response's advantage by comparing its reward to the average reward of a group of responses sampled for the same prompt.Mechanistic InterpretabilityMechanistic interpretability is the study of reverse-engineering the internal computations of neural networks into human-understandable algorithms. It aims to identify the features and circuits a model uses, so its behavior can be predicted, audited, and made safer.Differential PrivacyDifferential privacy is a mathematical definition of privacy that adds calibrated random noise to outputs so that the presence or absence of any single individual's data barely changes the result, bounded by a privacy budget epsilon.
Prompting & Reasoning
Prompt EngineeringPrompt engineering is the practice of designing and refining the text instructions given to a large language model so that its output is accurate, relevant, and formatted as intended. It treats wording, structure, examples, and constraints as levers that shape model behavior at inference time.Chain-of-Thought (CoT) PromptingChain-of-thought prompting is a technique that asks a large language model to generate intermediate reasoning steps before its final answer, improving accuracy on multi-step arithmetic, commonsense, and symbolic reasoning tasks.Zero-Shot PromptingZero-shot prompting is a technique where a large language model is asked to perform a task using only a natural-language instruction, with no worked examples or demonstrations included in the prompt. The model relies entirely on knowledge acquired during pretraining.System PromptA system prompt is the highest-authority instruction set given to a large language model that defines its identity, tone, rules, and output format before any user input arrives. It persists across a conversation and outranks ordinary user messages in the model's instruction hierarchy.Prompt InjectionPrompt injection is an attack in which untrusted input gets the language model to follow instructions the attacker embedded in that input, overriding the instructions the application's developer intended. It works because an LLM reads developer instructions and external data as one undifferentiated stream of tokens.Few-Shot LearningFew-shot learning is the ability to learn a new task or recognize a new class from only a handful of labeled examples. In modern large language models it most often means in-context learning: placing a few input-output demonstrations directly in the prompt so the model performs the task with no weight updates.In-Context LearningIn-context learning (ICL) is the ability of a large language model to perform a new task purely from instructions and example demonstrations placed in its prompt at inference time, with no changes to its weights. The model conditions on the prompt and predicts the next token; the apparent learning happens inside a single forward pass.Temperature and SamplingTemperature and sampling are the decoding controls that turn a language model's raw output scores into an actual token choice. Temperature scales the softmax to make the distribution sharper or flatter, while sampling methods like top-k and top-p decide which candidate tokens are eligible to be drawn.Context EngineeringContext engineering is the discipline of assembling everything that enters a language model's context window for a given step, including instructions, retrieved knowledge, tool outputs, memory, and conversation history, under a finite token budget. It treats the context as a state to curate rather than a single prompt to phrase.Self-Consistency (Decoding)Self-consistency is a decoding strategy that samples multiple chain-of-thought reasoning paths for the same prompt, then selects the final answer by majority vote. It improves reasoning accuracy by replacing a single greedy chain with an aggregate over diverse paths.Tree of Thoughts (ToT)Tree of Thoughts (ToT) is a reasoning framework that lets a language model explore multiple intermediate reasoning steps as a search tree, evaluating and backtracking among branches. It generalizes chain-of-thought from a single linear path to deliberate search.Structured Outputs (Constrained Decoding)Structured outputs are a feature that forces a language model's response to match a developer-supplied JSON Schema exactly, using constrained decoding to make invalid tokens impossible to sample.Test-Time Compute (Inference-Time Scaling)Test-time compute is the strategy of spending more computation during inference, by generating longer reasoning or sampling many candidate answers, to improve a model's accuracy without changing its weights.
Retrieval & Context
Embeddings (Text and Vector Embeddings)An embedding is a dense numeric vector that represents the meaning of a piece of data, such as text, an image, or audio, so that items with similar meaning land close together in vector space. This geometry lets machines compare meaning rather than match exact words.Vector Search (Similarity Retrieval)Vector search retrieves the items whose embeddings are closest to a query embedding in high-dimensional space, ranking results by a distance or similarity metric rather than by keyword overlap. Approximate nearest neighbor algorithms make this fast at scale by trading a small amount of recall for large speed gains.Vector DatabaseA vector database is a system that stores high-dimensional embeddings and serves fast approximate nearest neighbor queries over them, using index structures like HNSW or IVF plus metadata filtering, scaling, and persistence beyond what a bare similarity-search library provides.Semantic SearchSemantic search retrieves results by meaning rather than exact words. Queries and documents are converted into embedding vectors, and the system ranks candidates by vector similarity, so a query can match relevant text even when no keywords overlap.Retrieval-Augmented Generation (RAG)Retrieval-Augmented Generation (RAG) is a technique that grounds a language model's output in external knowledge: at query time the system retrieves relevant documents from a corpus and feeds them into the model's context so it generates answers based on fetched evidence rather than parametric memory alone.Context WindowA context window is the maximum amount of text, measured in tokens, that a language model can read and reference at one time when generating a response. It acts as the model's working memory: anything outside it is invisible unless re-supplied, summarized, or retrieved.RerankingReranking is a second retrieval stage that reorders an initial candidate list by scoring each query-document pair with a more accurate model, usually a cross-encoder, so the most relevant items rise to the top before they reach the generator.Hybrid SearchHybrid search is a retrieval strategy that runs both lexical (sparse, keyword-matching such as BM25) and semantic (dense, vector-similarity) search, then fuses the two ranked lists into one result set so each method covers the other's blind spots.Chunking StrategiesChunking strategies are the rules that split long documents into smaller passages before embedding and indexing them for retrieval. The strategy fixes the size, boundaries, and overlap of each chunk, which sets the ceiling on what a retrieval system can ever return.GraphRAGGraphRAG is a retrieval-augmented generation approach that builds an LLM-extracted knowledge graph of entities and relationships over a corpus, then answers questions by traversing that graph and summarizing graph communities, rather than retrieving only flat vector chunks. It targets multi-hop and global "sensemaking" questions that ordinary chunk-based RAG handles poorly.Approximate Nearest Neighbor (ANN)Approximate nearest neighbor (ANN) search finds vectors close to a query vector quickly by trading a small amount of accuracy for large speed gains, instead of scanning every vector exactly. It is the indexing layer that makes vector search at scale practical.Agentic RAGAgentic RAG is a retrieval-augmented generation design in which an AI agent, rather than a fixed pipeline, decides when, what, and how to retrieve. It runs a plan, retrieve, and critique loop that can rewrite queries, call multiple tools or sources, judge the evidence it gets back, and retrieve again before answering.HyDE (Hypothetical Document Embeddings)HyDE is a retrieval technique that asks a language model to write a hypothetical answer to a query, then embeds that answer and uses its vector to find real documents, improving zero-shot dense retrieval.BM25 (Okapi BM25)BM25 is a sparse keyword ranking function that scores documents for a query using term frequency, inverse document frequency, and document length normalization. It remains a strong baseline for full-text search.Reciprocal Rank Fusion (RRF)Reciprocal Rank Fusion is a method that combines several ranked result lists into one by summing the reciprocal of each item's rank in each list, fusing them without normalizing their underlying scores.Contextual RetrievalContextual Retrieval is a technique introduced by Anthropic that prepends a short, chunk-specific explanation to each document chunk before it is embedded and indexed, so retrieval keeps the surrounding context that naive chunking strips away. Anthropic reports it cuts retrieval failures by roughly a third, and by about two-thirds when combined with BM25 and reranking.Lost in the MiddleLost in the middle is a documented failure pattern in which language models use information placed at the beginning or end of a long context far more reliably than information placed in the middle. Named after a 2023 paper by Liu et al., it produces a U-shaped accuracy curve as a function of where the relevant content sits.HNSW (Hierarchical Navigable Small World)HNSW (Hierarchical Navigable Small World) is a graph-based algorithm for approximate nearest neighbor search over high-dimensional vectors. It builds a multi-layer proximity graph and uses greedy routing from a sparse top layer down to a dense base layer, giving sublinear search time with high recall, which makes it a default index in most vector databases.Vector Similarity Metrics (Cosine vs Dot Product vs Euclidean)Vector similarity metrics measure how close two embeddings are. Cosine similarity compares direction (angle), dot product combines direction and magnitude, and Euclidean distance measures straight-line distance. The right choice depends on whether embeddings are normalized.
Agents & Tools
AI Agents: Types and ArchitecturesAn AI agent is a system that perceives its environment, decides on actions toward a goal, and acts through effectors or tools. Modern LLM-based agents pair a language model with planning, memory, and tool use, and range from simple reflex designs to multi-agent architectures.Agentic AIAgentic AI is a design paradigm in which AI systems pursue goals autonomously by planning multi-step actions, using tools and memory, observing results, and adapting, rather than producing a single response to one prompt.Multi-Agent SystemsA multi-agent system is an architecture in which several LLM-based agents, each with a scoped role, tools, and context, coordinate to solve a task that one agent would handle less reliably alone. Agents are organized by a coordination pattern and exchange information through messages or shared memory.Function Calling (Tool Use)Function calling, also called tool use, is the capability that lets a large language model emit a structured request (a tool name plus JSON arguments matching a supplied schema) which the application executes and feeds back to the model. The model selects and parameterizes the call but never runs the code itself.Model Context Protocol (MCP)The Model Context Protocol (MCP) is an open, JSON-RPC 2.0 based standard, introduced by Anthropic in November 2024, that lets AI applications connect to external tools, data, and prompts through a uniform client-server interface, replacing bespoke per-integration connectors.ReAct AgentA ReAct agent is an LLM-driven loop that interleaves natural-language reasoning (Thought) with tool calls (Action) and tool results (Observation), repeating until it can answer. The pattern comes from the 2022 paper "ReAct: Synergizing Reasoning and Acting in Language Models."AI Agent MemoryAI agent memory is the system that lets an AI agent retain and reuse information beyond a single message. It splits into short-term memory (the context window, which acts as temporary working memory) and long-term memory (a persistent external store of facts, past events, and learned procedures that survives across sessions).ChatGPT ProjectsChatGPT Projects is a built-in feature that groups related chats, uploaded files, and project-specific custom instructions into a single workspace, and it can keep its memory scoped to that project so context does not leak in from or out to the rest of your account.AI Memory LayerAn AI memory layer is the persistence and retrieval tier that sits between an LLM's short-term context window and long-term personalization. It captures facts from conversations, stores them in a database (often vector plus graph), and injects the relevant pieces back into the prompt on later turns so an agent or app "remembers" a user across sessions.Agent OrchestrationAgent orchestration is the coordination of one or more LLM-driven agents, their tools, and shared state across a multi-step task. It covers how work is decomposed, routed, run in parallel or in sequence, and synthesized, and it spans both predefined workflows and more autonomous agent loops.Computer-Using Agent (Computer Use)A computer-using agent is an AI agent that operates software the way a person does: it views the screen through screenshots and acts through mouse and keyboard, rather than through dedicated APIs. It runs an agentic loop of seeing the screen, deciding an action, executing it, and taking a new screenshot.
Models & Evaluation
Model EvaluationModel evaluation (often called "evals") is the systematic measurement of a language model's capabilities, reliability, and cost using benchmarks, automated judges, and human preference comparisons, so that models can be scored, compared, and gated for release or production use.Reasoning Models (Thinking Models)Reasoning models, also called thinking models, are large language models trained to generate an extended internal chain of thought before answering. They spend extra inference-time (test-time) compute to plan, verify, and self-correct, trading speed and cost for higher accuracy on multi-step problems.Model Inference (LLM Serving)Model inference is the process of running a trained machine learning model on new inputs to produce outputs. For large language models, it means generating text one token at a time, and LLM serving is the engineering of doing this efficiently at scale.AI HallucinationAn AI hallucination is output from a generative model, especially a large language model, that is fluent and confident but factually wrong, unsupported, or inconsistent with its source, such as an invented statistic or a fabricated citation.Explainable AI (XAI)Explainable AI (XAI) is the set of methods and design practices that make an AI system's outputs understandable to humans, so that people can comprehend, appropriately trust, and effectively manage the model's decisions rather than treating it as an opaque black box.GPT-4 and GPT-4o: OpenAI's Flagship Model FamilyGPT-4 and GPT-4o are large multimodal models from OpenAI. GPT-4 (2023) accepted text and images and returned text; GPT-4o (2024) was the "omni" model trained natively across text, vision, and audio with a 128K-token context. As of 2026 the GPT-5 family had succeeded them as ChatGPT's default.Claude: Anthropic's Model Family (Opus, Sonnet, Haiku)Claude is a family of large language models built by Anthropic, organized into three tiers: Opus (most capable), Sonnet (balanced speed and intelligence), and Haiku (fastest and lowest cost). All tiers support text, vision, and tool use, and are trained with Constitutional AI.Llama: Meta's Open-Weight Model FamilyLlama is Meta's family of open-weight large language models, distributed with downloadable parameters under a custom community license. It spans Llama 1 through the Llama 4 herd (Scout, Maverick, Behemoth) and is widely used for self-hosting and fine-tuning.Grok: xAI's Model FamilyGrok is the family of large language models built by xAI, Elon Musk's AI company. The lineage runs from the open-weight Grok-1 (2023) to the Grok 4.x reasoning flagships (as of 2026), and is distinguished by long-context reasoning, agentic tool use, and tight integration with the X (Twitter) platform for real-time data.Adversarial Attacks (AI)Adversarial attacks are inputs deliberately perturbed to make a machine learning model misclassify or misbehave, often with changes too small for a human to notice. They expose a basic fragility in how neural networks map inputs to outputs.KV CacheA KV cache stores the Key and Value tensors computed for previous tokens during autoregressive decoding, so each new token reuses past attention work instead of recomputing it over the whole sequence.Stable DiffusionStable Diffusion is an open-weights text-to-image latent diffusion model, first released in August 2022, that generates images by denoising a compressed latent representation conditioned on a text prompt. Operating in a low-dimensional latent space rather than on raw pixels lets it run on consumer GPUs.BERTBERT (Bidirectional Encoder Representations from Transformers) is an encoder-only Transformer model from Google, introduced in 2018, that learns deeply bidirectional contextual representations of text by jointly conditioning on both left and right context. It is pretrained with masked language modeling and then fine-tuned for understanding tasks like classification, question answering, and named entity recognition.Small Language Models (SLM)Small language models (SLMs) are compact transformer-based language models, loosely those under roughly 10 billion parameters, designed to run cheaply, on-device, or at low latency while retaining strong performance on many tasks.DeepSeek ArchitectureDeepSeek is a Chinese AI lab whose open-weight models combine Multi-head Latent Attention (MLA) for a compressed KV cache, a fine-grained DeepSeekMoE mixture-of-experts design, and reinforcement-learning-trained reasoning, reaching frontier quality at low reported training compute.Edge AIEdge AI runs machine learning inference directly on the device that captures the data, such as a phone, laptop, camera, car, or microcontroller, instead of sending the data to a remote cloud server. It trades raw model size for low latency, offline operation, and keeping data local.Emergent CapabilitiesEmergent capabilities are abilities that large language models appear to gain abruptly at larger scale, absent in smaller models. Whether these jumps are real or an artifact of the chosen evaluation metric is actively debated.Gemini 1.5 ProGemini 1.5 Pro was a natively multimodal, sparse mixture-of-experts model from Google DeepMind, announced in February 2024 and notable for a context window of up to 1 million tokens (later 2 million). It anchored the Gemini family, which by 2026 had advanced through the 2.x and 3.x generations.Neural Scaling LawsNeural scaling laws are the empirical finding that a language model's test loss falls as a smooth power law in three quantities: the number of parameters, the size of the training dataset, and the compute spent. They let researchers predict the performance of large models from cheap small-scale runs.Speculative DecodingSpeculative decoding is an inference-acceleration method that speeds up autoregressive LLM generation without changing the output distribution: a small fast draft model proposes several future tokens, the large target model verifies them in one parallel forward pass, and a rejection-sampling rule keeps the longest correct prefix.World ModelsA world model is a learned internal model of an environment's dynamics that lets an agent predict how the world will change and what its actions will do, so it can plan or imagine rollouts instead of acting purely reactively.LLM-as-a-JudgeLLM-as-a-judge is the practice of using a strong language model to grade or compare other models' outputs in place of human raters, scoring responses pointwise, pairwise, or against a reference. It trades human cost and latency for scalable, automatable evaluation that correlates reasonably well with human preference but carries its own biases.RAG EvaluationRAG evaluation is the practice of measuring a retrieval-augmented generation system by scoring its retrieval stage and its generation stage separately, using metrics such as recall@k and context relevance for retrieval and faithfulness and answer relevance for generation.PerplexityPerplexity is an intrinsic metric for language models that measures how well a model predicts a sample of text. It equals the exponential of the average per-token cross-entropy, and can be read as the effective number of equally likely choices the model considers at each step. Lower perplexity means better prediction.CLIP (Contrastive Language-Image Pre-training)CLIP is a model from OpenAI that trains an image encoder and a text encoder together so matching image-caption pairs land near each other in a shared embedding space, enabling zero-shot image classification from text labels.Data Drift (and Concept Drift)Data drift is a change in the statistical distribution of a model's input features over time, while concept drift is a change in the relationship between inputs and the target. Both degrade a deployed model's accuracy and must be monitored in production.Precision, Recall, and F1 ScorePrecision is the fraction of predicted positives that are correct, recall is the fraction of actual positives that are found, and F1 is their harmonic mean. They evaluate classifiers, especially on imbalanced data where accuracy misleads.
AI & Retrieval
Optical character recognition (OCR)Optical character recognition (OCR) is technology that converts images of typed, printed, or handwritten text (scans, photos, or image-only PDFs) into machine-encoded text that a computer can edit, search, and store.Intelligent document processing (IDP)Intelligent document processing (IDP) is an AI-driven approach to capturing, classifying, extracting, and validating data from documents, combining optical character recognition with natural language processing and machine learning to interpret content rather than just read characters.Knowledge graphA knowledge graph is a structured representation of information that stores real-world entities as nodes and the relationships between them as edges, forming a network of machine-readable facts that computers can query, traverse, and reason over.Episodic, Semantic and Procedural Memory (in AI)Episodic, semantic, and procedural memory are three categories of human long-term memory (events, facts, and skills) that AI researchers borrow to organize how agents store experience, knowledge, and learned behavior. Episodic memory recalls specific events, semantic memory holds general facts, and procedural memory encodes how to do things.
Memory & Cognition
Forgetting curveThe forgetting curve is the roughly exponential decline in retention of newly learned information over time when no effort is made to review it, first measured by Hermann Ebbinghaus in 1885.Spaced repetitionSpaced repetition is a learning technique that schedules reviews of material at increasing intervals over time, exploiting the spacing effect to move knowledge into durable long-term memory with far less total study than cramming.Cognitive offloadingCognitive offloading is the use of physical action or the external environment (writing notes, setting reminders, using GPS or search) to reduce the information-processing demands placed on the brain, shifting mental effort from internal memory to external tools.Context RotContext rot is the measurable decline in a language model's output quality as its input context grows longer, even while the prompt stays well under the model's hard token limit. It describes degraded use of a filled window, not the size of the window itself.ChatGPT MemoryChatGPT Memory is OpenAI's personalization system that lets ChatGPT carry information about you across separate conversations. It combines saved memories you can view and edit with a "reference chat history" layer that recalls details from past chats. As of mid-2026 it is generated by a background synthesis process OpenAI calls Dreaming.Claude MemoryClaude Memory is Anthropic's feature that lets Claude retain context across separate conversations by building an editable, project-scoped summary of your preferences and ongoing work, complemented by a "search past chats" tool on paid plans. It is optional, controllable from Settings, and can be paused, reset, or bypassed with incognito chats.Gemini PersonalizationGemini personalization is Google's set of features that let the Gemini app tailor responses to an individual user by drawing on three sources: the memory of past Gemini chats ("Personal context"), saved preferences and custom instructions, and, with opt-in connections, content and activity from Google services like Gmail, Photos, Search, and YouTube ("Personal Intelligence"). The past-chats memory layer is on by default for eligible personal accounts, while connecting Google apps is opt-in.Reference Chat HistoryReference chat history is the ChatGPT memory setting that lets the model draw on information from your past conversations automatically, without you asking it to remember anything. It is the implicit cross-chat layer that sits alongside the editable "Saved memories" list and can be toggled on or off under Settings > Personalization > Memory.MemGPT (LLM as Operating System)MemGPT is a system that treats an LLM like an operating system, using virtual context management to page information between a limited in-context main memory and unlimited external storage, giving the model effectively unbounded, self-managed memory.Generative Agents (Memory Stream & Reflection)Generative Agents are LLM-driven simulated characters from Stanford and Google research that store experiences in a natural-language memory stream, score memories by recency, importance, and relevance for retrieval, and synthesize higher-level reflections to drive believable behavior.Working Memory vs Long-Term Memory (in AI Agents)In AI agents, working memory is the small, fast, temporary information held in the context window during a task, while long-term memory is durable knowledge stored externally (in databases or vector stores) and retrieved across sessions. Most capable agents combine both.
PKM & Note-taking
Second brainA second brain is an external, digital system for capturing, organizing, and retrieving the information, ideas, and references a person encounters, designed to extend biological memory and support thinking. The term was popularized by Tiago Forte.Personal knowledge management (PKM)Personal knowledge management (PKM) is the ongoing practice of capturing, organizing, distilling, retrieving, and sharing the information an individual encounters, so that scattered notes and sources become a usable personal knowledge base rather than a forgotten archive.ZettelkastenThe Zettelkasten (German for "slip box") is a note-taking and knowledge-management method built from many small, atomic notes, each given a unique identifier and explicitly linked to related notes, forming a navigable web of ideas rather than a filed archive.