TerranSoul // Brain Architecture

GIVING AN LLM A BRAIN

The model alone forgets everything between turns. We graft on the regions of a real brain — a frontal lobe to focus, a hippocampus to remember, a cerebellum for skill. Hover a region to see how it works.

Read the architecture design doc →
Read the benchmark comparison →
MCP · brainstem · the one nerve
{{ lobe.title }}
{{ lobe.region }}
{{ ch.title }}
LLM Core
neocortex
{{ panelKicker }}
{{ panelTitle }}
{{ panelPart }}
{{ panelPlain }}
faculties {{ panelFaculties }}
{{ m.label }}
prompt → LLM → reply no memory between turns turn gate SKIP · 0 cost LIGHT NORMAL FULL pipeline you Lyon France café EU 2 hops: you → Lyon → France vector bm25 graph fused RRF blends 3 rankings → 1 list k=60 = smoothing, so no method dominates SHORT · last ~20 turns WORKING · this session LONG · durable, decays used again = promoted · ignored = fades reply text "that's wonderful!" sentiment → avatar expression skill: "reset the cache" confidence 0.67 → rising promote L4 → L1 3 fails in a row → quarantine QUERIED BY PULLS FROM AGENTS → Claude Code Codex OpenClaw Hermes TerranSoul (peer) APPS → Desktop app Mobile app MCP one nerve brain TerranSoul Connectors repo brains peer TerranSoul RSS · GitHub · IMAP · FS over MCP / LAN one MCP nerve → many consumers · fed by its own sources client BRAIN a sense feeding the same brain
{{ tag }}
{{ panelTech }}