Glossary
AI agent memory glossary
Key terms and concepts behind AI agent memory systems, cognitive architectures, and adaptive retrieval.
Adaptive Strategy Selection
The process of automatically choosing the best retrieval approach based on historical performance. MetaMemory uses multi-armed bandit algorithms (Thompson Sampling and Upper Confidence Bound) to learn which of its 5 retrieval channels works best for each type of query, continuously optimizing based on real usage patterns.
BYOK (Bring Your Own Keys)
An architecture where users provide their own API keys for external services rather than using shared credentials. MetaMemory's BYOK model means embeddings are generated using your own provider credentials — OpenAI, Cohere, Voyage AI, etc. MetaMemory never sees your API keys or raw data, and you maintain full control over your embedding pipeline.
Drift Detection
Statistical monitoring of model performance to detect when the underlying data distribution shifts. When usage patterns change (new topics, different user behavior), MetaMemory detects the drift within approximately 50 queries and accelerates adaptation to the new distribution, preventing stale retrieval strategies from degrading quality.
Embedding Provider
A service that converts text into vector representations (embeddings) that capture semantic meaning. MetaMemory supports multiple embedding providers including OpenAI, Google Gemini, Cohere, Voyage AI, Mistral, and Ollama. Each provider has different strengths — Cohere excels at retrieval, Voyage AI at domain-specific content, Ollama at privacy-first deployments.
Emotional Intelligence (AI)
The capability of an AI system to detect, understand, and respond to emotional signals in user interactions. MetaMemory detects 6 computational emotional states in real time through analysis of linguistic markers, interaction patterns, and conversational dynamics — enabling agents to adapt their tone and approach based on how a user feels.
Emotional Memory
The affective dimension of memory that captures how an experience felt, not just what happened. MetaMemory detects and encodes 6 computational emotional states (confident, uncertain, confused, frustrated, insight, breakthrough) in real time, allowing agents to adapt their behavior based on emotional context across sessions.
Episodic Memory
A cognitive science concept referring to memory of personal experiences and specific events, including their temporal and contextual details. Unlike semantic memory which stores general facts, episodic memory encodes when something happened, what preceded it, and what followed. MetaMemory draws inspiration from this concept through its "context" embeddings, which capture situational metadata and temporal relationships to enable experience-like retrieval across sessions.
Importance Weighting
A mechanism for scoring the relative significance of memories based on multiple signals. During consolidation, MetaMemory evaluates importance using recency, frequency, emotional weight, and downstream utility. High-importance memories are preserved in full while routine interactions are aggressively compressed.
Memory Consolidation
The process of merging, compressing, and strengthening memories over time — analogous to what happens during sleep in the human brain. MetaMemory uses LLM-powered consolidation to merge related memories, compress redundant information, and strengthen important connections, achieving 70% compression while preserving 97% recall quality.
Multi-Vector Embeddings
An encoding approach that represents each memory across multiple distinct vector spaces simultaneously. Traditional memory systems use a single embedding per memory, losing nuance. MetaMemory encodes every interaction across four vector spaces — semantic, emotional, process, and context — capturing the full dimensionality of each memory for more precise retrieval.
Online Learning
A machine learning paradigm where the model updates continuously as new data arrives, rather than being retrained in batches. MetaMemory's retrieval models adapt in real time after every interaction, with drift detection to handle changing usage patterns and automatic rollback if an adaptation degrades performance.
Procedural Memory
A cognitive science concept referring to memory for skills, habits, and how-to knowledge — actions you perform automatically without conscious recall. MetaMemory draws inspiration from this concept through its "process" embeddings, which capture step-by-step workflows, procedures, and task sequences, enabling task-oriented retrieval — such as remembering how a user prefers to deploy code.
Reciprocal Rank Fusion (RRF)
A technique for combining ranked results from multiple retrieval channels into a single unified ranking. RRF assigns scores based on rank position rather than raw similarity scores, making it robust to differences in score distributions across channels. MetaMemory uses RRF to fuse results from its 5 retrieval channels into a single coherent result set.
Semantic Memory
In cognitive science, a type of long-term memory that stores general knowledge and facts independent of personal experience. In MetaMemory, the semantic embedding type encodes the factual meaning of interactions — what was discussed, what decisions were made, what information was shared — powering fact-based recall and knowledge queries as one of four vector types (semantic, emotional, process, context).
Temporal Context
Metadata that captures when a memory was formed and its position in a sequence of events. Each memory in MetaMemory carries temporal information: when it happened, what preceded it, what followed. This allows agents to reason about sequences, durations, and the order of events — critical for context-based retrieval and cross-session continuity.
Thompson Sampling
A Bayesian algorithm for the multi-armed bandit problem that maintains probability distributions over the effectiveness of each option. In MetaMemory, Thompson Sampling selects which of the 5 retrieval channels to prioritize for each query, naturally balancing exploration of less-tested channels with exploitation of known-good ones. It converges on optimal strategies within approximately 200 queries.