Cognitive Landscape

An experimental cognitive layer for studying how experience changes a system itself, rather than what it can look up. It is not a memory retrieval system: nothing the user says is stored, and there is no index to search.

The question it tries to make testable: can a history of interaction leave persistent internal structure that changes what can happen later?

Core mechanism

Every structure holds three states:

| State | Role | |---|---| | m | fast state; reacts first to local mismatch, decays with real elapsed time | | theta | slow structure; changes only by consolidating a fraction of m | | p | plasticity; gates how much of m reaches theta, and is itself changed by instability |

One local update (landscape/plasticity.py) is: predict with theta + m, measure mismatch against the input, move m toward the mismatch, consolidate p * rate * m into theta, keep the rest in m, then raise p when |m| stays large and lower it when things are quiet. There is no reward, no importance score, no central judge, and no LLM deciding what matters.

What this version does (v0.2)

Two perception channels feed the same equations:

1. Embedding channel (primary). Each turn of text is embedded locally (BAAI/bge-small-zh-v1.5 through fastembed/ONNX Runtime, no torch) and routed by cosine similarity into unnamed vector structures, where theta is a region centre, m a fast offset and p a scalar. The v0.1 scalar equations are applied unchanged to vectors via a minimal Vector type (landscape/vector.py). An input far from every region starts a new one; thresholds and the measurements behind them are documented in VectorLandscapeConfig (landscape/config.py). The model loads on a background thread, so a host never blocks on it.