Home / Documentation / Configuration Guide

Memory System

Continuous memory — companionship that never forgets

Neurova agents have genuine persistent memory, not a context that is wiped after every conversation.

Memory tiers

TierDescriptionCharacteristics
Working memoryThe current conversation contextUpdated in real time, configurable capacity
Short-term memoryRecent events and conversationsTemperature decay, automatic discovery of hot memories
Long-term memoryPersisted experience and knowledgeConsolidated storage, version control, retrospective access
Emotional memoryMemories tagged with affectDriven by the emotion engine, steering reasoning and decisions
Temporal knowledge graphA graph of events and entitiesSupports causal reasoning and associative retrieval

The 17-dimension memory taxonomy

Where conventional agent frameworks only distinguish "short-term" from "long-term", Neurova partitions memory into 17 types: conversation, fact, profile, relationship, skill, experience, lesson, emotion, goal, plan, project, event, insight, pattern, preference, config and reflection.

NeRF-enhanced memory retrieval (six-channel volumetric fusion)

Neurova transfers the volumetric rendering theory of NeRF (Neural Radiance Fields) into memory retrieval, shifting the paradigm from discrete search to rendering a continuous semantic field.

The six channels

ChannelFunctionDefault density
TemperatureRetrieval by memory heat0.7
TextSemantic similarity0.9
CategoryMemory type matching0.5
GraphKnowledge-graph association0.6
EmotionEmotional state matching0.8
VoiceVoiceprint matching0.4

Fusion modes

  • Legacy mode: weighted sum, score = Σ(w_i · c_i)
  • NeRF mode: volumetric fusion, score = Σ(T_i · σ_i · c_i · w_i), with transmittance decaying across channels

The memory temperature mechanism

  • Temperature engine: Ebbinghaus forgetting curve + Bayesian forgetting + consolidation
  • Sleep processing: five sleep stages + an idle consolidation chain (idle_tracker) + dream generation
  • Muscle memory: three promotion tiers — L1 exact keywords / L2 vector similarity / L3 tool memory

Sleep and resource management

Neurova simulates a human sleep-wake cycle: awake (Beta 13-30Hz) / light sleep (Alpha 8-13Hz) / deep deliberation (Theta 4-8Hz) / REM (Gamma 30+ Hz), with automatic sleep, manual control and scheduled wake-up.