Hierarchical multi-layer action store for agent-based task execution

1 . An apparatus comprising:

a processor;

a first memory comprising an action store, wherein the action store comprises a hierarchical arrangement of memory layers and a plurality of actions stored in one or more of the memory layers, and the action store stores first mappings among the plurality of actions and second mappings among the plurality of actions and one or more of the memory layers, wherein the first mappings among the plurality of actions are hierarchical and an action of the plurality of actions is executable by an agent; and

a second memory comprising an instruction to cause the processor to update the first mappings or the second mappings in response to a signal received from a device, the signal relating to (i) a requested task, (ii) an execution of the action by the agent, (iii) performance data associated with the action, or any combination of (i), (ii), (iii);

wherein the memory layers comprise a first layer storing a first action and a second layer storing a second action, wherein a first security policy is associated with the first layer, a second security policy, different from the first security policy, is associated with the second layer, and wherein the instruction is to cause the processor to control access to the first layer according to the first security policy and control access to the second layer according to the second security policy.

2 . The apparatus of claim 1 , wherein the action comprises metadata and the metadata comprises an indication of a security level associated with the action, or a performance metric associated with the action, or the indication of the security level and the performance metric, and wherein the instruction is to control access to the action according to the indication of the security level and a rule, or the performance metric and the rule, or the security level, the performance metric, and the rule.

3 . The apparatus of claim 1 , wherein the first memory is indexed by computing an embedding of the action and mapping the embedding to the action stored in the action store.

4 . The apparatus of claim 1 , wherein the instruction is to cause the processor to (i) move the first action from the first layer to the second layer, (ii) move the second action from the second layer to the first layer, (iii) combine a plurality of actions into the first action, (iv) expand the first action into a plurality of actions, or any combination of (i), (ii), (iii), (iv).

5 . A computer-implemented method comprising:

using an input to determine a task, wherein the input is obtained via a device and an agent accessible via the device;

using the task to search an action store, wherein the action store comprises a hierarchical arrangement of memory layers and a plurality of actions stored in one or more of the memory layers, wherein the memory layers comprise a first layer storing a first action and a second layer storing a second action, wherein a first security policy is associated with the first layer, a second security policy, different from the first security policy, is associated with the second layer, and wherein an instruction is to cause a processor to control access to the first layer according to the first security policy and control access to the second layer according to the second security policy;

in response to determining that the action store comprises an action that corresponds to the task, providing access to the action to the agent; and

triggering the agent to execute the action.

6 . The method of claim 5 , further comprising:

providing the input and a first instruction to a machine learning model, wherein the first instruction is to cause the machine learning model to determine the task; and

receiving the task from the machine learning model in response to the first instruction.

7 . The method of claim 6 , further comprising:

providing the task and a second instruction to the machine learning model, wherein the second instruction is to cause the machine learning model to identify the action that corresponds to the task using embedding-based retrieval, retrieval-augmented generation, or embedding-based retrieval and retrieval-augmented generation; and

receiving the identified action from the machine learning model in response to the second instruction.

8 . The method of claim 7 , wherein the second instruction is to instruct the machine learning model to determine requirements for performing the task and to determine whether the action is capable of performing the requirements.

9 . The method of claim 8 , further comprising:

in response to determining that the action is not capable of performing the requirements, providing the task and a third instruction to the machine learning model, wherein the third instruction is to cause the machine learning model to iteratively decompose the task into sub-tasks until the action is capable of performing the sub-tasks.

10 . The method of claim 6 , further comprising:

providing the task and a fourth instruction to the machine learning model, wherein the fourth instruction is to cause the machine learning model to identify a plurality of actions that correspond to the task, rank the identified plurality of actions according to a first criterion associated with the task, and select an action from the ranked plurality of actions according to a second criterion associated with the task; and

receiving the selected action from the machine learning model in response to the fourth instruction.

11 . The method of claim 6 , further comprising:

in response to determining that a value of a performance metric associated with a first version of an action stored in the first layer meets or exceeds a performance criterion, providing access to the agent to execute the first version of the action from the first layer.

12 . The method of claim 6 , wherein the agent is a robot or self-driving vehicle.

13 . The method of claim 6 , further comprising:

determining a mutability characteristic of the action, wherein the mutability characteristic is an indication of ability of the agent to self-mutate by executing the action; and

using the mutability characteristic to control the access to the action by the agent.

14 . A computer-implemented method comprising:

receiving a signal via an agent, wherein the signal relates to a first execution of a task by the agent using an action store, wherein the action store comprises a hierarchical arrangement of memory layers, and the action store comprises a plurality of actions stored in one or more of the memory layers, and the action store stores first mappings between actions and second mappings between the plurality of actions and the one or more of the memory layers, wherein the first mappings among actions are hierarchical, and an action is executable by the agent, wherein the memory layers comprise a first layer storing a first action and a second layer storing a second action, wherein a first security policy is associated with the first layer, a second security policy, different from the first security policy, is associated with the second layer, and wherein an instruction is to cause a processor to control access to the first layer according to the first security policy and control access to the second layer according to the second security policy;

using the signal to modify the action store; and

providing a modified portion of the action store to a second execution of the task by the agent.

15 . The method of claim 14 , further comprising:

using the signal to create or update the first mappings, the second mappings, or the first mappings and the second mappings.

16 . The method of claim 14 , further comprising:

in response to the signal, (i) moving the first action from the first layer to the second layer of the memory layers, (ii) moving the second action from the second layer to the first layer, or any combination of (i) and (ii).

17 . The method of claim 14 , further comprising:

using the signal and a trajectory optimization mechanism to combine a plurality of actions into a third action or expand the third action into the plurality of actions.

18 . The method of claim 14 , wherein the task comprises detecting and blocking malicious access to a cloud service and the agent is security software executing on a user computing device.

19 . The method of claim 14 , wherein the agent is a robot or self-driving vehicle.