ECHO-Core

Intelligence that carries its experience forward.

ECHO-Core brings together echoai’s development through its 5.9 research line, focused on autonomous robotics and CPU-based intelligence. Its core preserves experience, recalls consequences and uses shared memory to help agents cooperate.

WSP represents evidence in 16-byte packets. Memory distinguishes direct observations from reports by other agents. The core can operate without a language model: experience remains useful even without a GPU.

Code structure

ECHO-Core structure: environment, agent, consequence memory, behaviour rules and WSP-based cooperation. — echoai modules used by MEMORIA-X and ENJAMBRE-1. Arrows show dependencies and information exchange. Enlarge diagram

Learn in one world, use it in another

MEMORIA-X transfers memory of consequences to new environment families. In the A1 evaluation family, mean harm falls from 150.0 with empty memory to 24.2 with learned memory.

Mean harm in MEMORIA-X: 150 with empty memory versus 24.2 with learned memory.

Improve round after round

ACUMULA-3 selects and combines available modifications based on measured failures. Across five rounds, the mean falls from 94.3 to 13.8 failed episodes out of 120: five consecutive improvements.

Solve tasks together

ENJAMBRE-1 connects four agents with private and shared global memory. The team solves all 160 evaluation worlds in 532 episodes; a single agent solves 119 in 1,471 episodes.

ENJAMBRE-1 aggregate cost: 920 without global memory versus 691 with shared memory.

Results from MEMORIA-X, ACUMULA-3 and ENJAMBRE-1, validated in ECHO-5.8 and carried forward in echoai’s development towards ECHO-Core. These evaluations use simulated environments.