OpenPsi

help edit space_dashboard
Draft — This content has not been approved for publication.
ai generated
open_in_full page
fullscreen modal
edit edit
space_dashboard advanced

Home / Implementation Families / Attention and Motivation / OpenPsi

OpenPsi is the classical motivation and action-selection framework — a computational realization of Dietrich Dörner's Psi theory of cognition, in which behaviour is driven by competing demands (urges) that are shaped by emotional modulators and discharged by selecting context-appropriate actions. It answers "what should the agent want, and what should it do next?", complementing ECAN, which answers "what should the agent pay attention to?". OpenPsi is the heuristic predecessor that MetaMo later re-casts in formal category theory.

Last verified: 2026-06-08

On this page

Lineage

OpenPsi descends from a specific line of motivational-cognition research:

  • Psi theory (Dietrich Dörner). A psychological theory in which cognition is organized around a small set of basic drives/demands and a handful of modulating parameters (such as arousal and resolution level), rather than around explicit goals alone.
  • MicroPsi (Joscha Bach). The first computational operationalization of Psi theory, giving the demands and modulators concrete dynamics.
  • OpenPsi (Cai, Goertzel, Geisweiller and collaborators, ~2011). A re-implementation of the MicroPsi/Psi ideas natively inside the classical AtomSpace, so that motivation could drive — and be driven by — the rest of the cognitive stack.
  • MetaMo (2025). The next-generation successor that formalizes the appraisal/decision cycle as a category-theoretic pseudo-bimonad; see MetaMo.

The practical evolutionary path inside Hyperon is OpenPsi (heuristic) → hyperon-openpsi (MeTTa port) → MetaMo (formal), with all three coexisting during the transition.

Core mechanics

  • Demands (urges). Scalar "needs" the agent tries to keep satisfied (for example energy, competence, affiliation). Each demand has a satisfaction level that decays over time and is replenished by appropriate actions.
  • Modulators. A small set of parameters (such as arousal, resolution level, and goal-directedness) that shift how the agent appraises situations and selects behaviour — the "emotional" tuning of cognition.
  • OpenPsi rules. Behaviour is encoded as Context → Procedure → Goal rules: when a context holds, running a procedure is expected to satisfy a goal (and thereby a demand). An action-selection step picks among the applicable rules based on demand urgency and modulator state.

Implementations

  • Classic OpenPsi (Scheme, legacy) — the original AtomSpace implementation, including the action-selector and a dynamics/ module built out through 2017. Now historical.
  • hyperon-openpsi (MeTTa + Python, operational) — Dörner's Psi theory ported to MeTTa: demands, emotions, modulators, and hill-climbing action planners. Demonstrated on ping-pong, a curious-agent (LLM-driven), and a Qwestor API use case. This is the operational predecessor substrate for MetaMo.
  • PeTTa-OpenPSI (MeTTa + Python, research/operational) — a parallel port on Prolog-native MeTTa (PeTTa) with A* and Thompson-sampling planners and a NARS-based trading use case.

The two MeTTa ports implement the same framework with different planners; consolidating them is an open opportunity noted on the Attention and Motivation family card.

Relationship to attention and to MetaMo

OpenPsi and ECAN are complementary: OpenPsi decides what to do (goal and action selection) while ECAN decides what to think about (attention allocation). Historically the two were briefly coupled — an OpenPsi default action-selector could bias attention by writing Short-Term Importance (STI) — but that path, like the other narrow ECAN coupling hooks, was deliberately decoupled by 2019 and has no MeTTa equivalent today. OpenPsi is best understood as an ECAN-coupling enabler rather than a default attention consumer; the full reconstructed coupling timeline lives on ECAN Full → Development and Historical Context. MetaMo is the formal successor — it keeps OpenPsi's demands-and-modulators intuition but replaces the hand-tuned dynamics with a category-theoretic appraisal/decision cycle.

Historical: the Hanson / Sophia dialogue branch

Outside the research line above, Hanson Robotics' Sophia stack consumed Classic OpenPsi as a behaviour-loop authority for dialogue and motor schemas (via ros-behavior-scripting and loving-ai-ghost), with a psi-step loop and psi-rule registrations mapping perceptions to behaviours. This was a runtime/application branch rather than a faithful reimplementation — demands were satisfied by hardcoded truth values rather than OpenPsi's quantitative satisfaction economy, and the production Loving AI Ghost runtime explicitly disabled attention weighting in 2018. It has no current Hyperon MeTTa equivalent.

Sources