Research & development

Memnon is an experimental memory architecture that translates the operating principles of human memory into a software environment. The system is built on the following cognitive-science principles: the Ebbinghaus forgetting curve, Collins & Loftus spreading activation, Diekelmann & Born sleep-based consolidation, Brown & Kulik flashbulb-memory formation, and 10 further principles.

The system runs on a tick-based scheduler (10-second cycle, circadian rhythm). During the sleep phases, NREM consolidation (daily summary, contradiction handling) and REM dreaming (association generation) take place. Memories are stored in an associative memory graph: weighted, typed edges, differentiated decay.

Dual-LLM architecture: a fast chain (gpt-oss-20b, System 1 — fast, intuitive) and a quality chain (llama-3.3-70b, System 2 — slow, deliberative). Multi-provider fallback chain: Groq → Google Gemini → local Ollama (NVIDIA Jetson AGX Orin 64GB).

Emergent phenomena: the “tip-of-the-tongue” state, spontaneous dream associations, flashbulb-memory formation — all of which emerged from the rules of the architecture, not from explicit programming.

Neuroanatomical mapping: the architecture can be functionally mapped to the brain's memory-related regions — the roles of the thalamus (input filtering), the hippocampus (episodic encoding), the prefrontal cortex (decision-making) and the neocortex (consolidated knowledge) are performed by software modules.

Camera: ZED X Mini (GMSL interface) — stereo depth sensing, visual odometry. The Memnon memory system will serve as the robot's cognitive memory: persistent storage of episodic memories, environmental context and learned behavioural patterns.