The description calls itself a performance optimization system for the harness rather than for the agent. That distinction is the whole product: every lever it names sits outside the model.
What the harness can be taught to do without retraining anything.
The unusual one. No other project on today's board uses this word.
The third repo on this board this week to put memory in its own one-line pitch.
Named alongside capability rather than after it, which is a scheduling decision.
A development order, not a feature. It describes how work gets sequenced.
Skills are written down and memory is stored. An instinct is neither: it implies behaviour the harness applies without being asked each time. That is a real claim about where judgement lives, and it is the line worth reading the repository to check.
No benchmark, no latency, no token figure, no before-and-after. A project with optimization in its own name invites those numbers, and the trending description publishes none, so this page prints none.
Concept mock, not affiliated with the project. The five concerns and four host names above are quoted in order from the repository's own GitHub Trending description on 24 August 2026: "The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond."