v0.9.0 — Neuroscience-Grounded Cognitive Architecture

[idapixl]

What's in v0.9.0

This release implements 16 cognitive improvements derived from three expert analyses: computational neuroscience, CS/ML theory, and information geometry.

The core insight: cortex-engine's mechanisms aren't biological metaphors — they implement the same mathematics. FSRS matches biological forgetting curves because the optimization problem is identical. Prediction error gating matches dopaminergic signaling because it's the optimal solution to "what should I remember?" Taking this seriously means treating biology as a specification, not an inspiration.

The Big Additions

Two-phase dream — dreamPhaseA() (NREM: compress) + dreamPhaseB() (REM: integrate). Run compression during sessions, integration in cron. Mirrors biological NREM→REM cycling.

GOAL nodes — goal_set stores desired future states that generate forward prediction error. This is the VTA/dopamine value channel — the missing half of the PE system. Goals bias what gets consolidated and what gets explored.

Neuroscience-grounded retrieval — Query-conditioned spreading activation (Synapse paper), GNN neighborhood aggregation, multi-anchor Thousand Brains voting.

Graph health metrics — Fiedler value (algebraic connectivity) measures knowledge integration. PE saturation detection prevents identity schema ossification.

Full Changelog

See the release notes for the complete list.

New Wiki Page

The Cognitive Architecture wiki page documents the full architecture with references.

Install

npm install cortex-engine@0.9.0

27 MCP tools. Zero breaking changes. 1,275 lines added.