Research Map

Octopus uses these papers as pressure tests for architecture, not as templates to clone.

Biological Octopus

Xitong Liang, Peking University: cephalopods as a model for alternative AI and robotics algorithms because color change and soft-arm control both require high-dimensional motor output. Octopus maps this to one intent layer plus local nervous systems at the tool edge.
The dynamics of pattern matching in camouflaging cuttlefish: flexible, exploratory pattern matching rather than one fixed component sequence. Octopus maps this to stable Needs with replaceable implementations.
Neuronal segmentation in cephalopod arms: segmented axial nerve cords and suckerotopy. Octopus maps this to tentacles with internal modules, local feedback, and compact Feed.

ReAct: reasoning and environment actions. Octopus moves concrete action work into tentacles.
Toolformer: learned tool-use decisions. Octopus makes tool choice a harness/tentacle concern.
Gorilla: accurate API use over changing docs. Octopus treats tools as intelligent surfaces.
ToolLLM: large-scale tool learning. Octopus keeps the kernel small so many tool systems can attach.

Reflexion: feedback and memory improve later trials. Octopus persists memory and route scores.
Voyager: executable skill growth from environment feedback. Octopus separates skill growth from the clean brain.

MemGPT: context as managed memory. Octopus keeps memory as a harness-evolved beat.
Self-RAG: demand-driven retrieval and critique. Octopus expresses demand as Need and feeds evidence through Feedback.

DataDriven cognitive supply: once the brain emits a need, the harness should learn how to satisfy it with the best available feed.