Multi-agent artificial intelligence system with shared experience repository

A method, comprising:

• maintaining, by one or more processors, a database storing a plurality of data records accessible to a language model, each data record comprising data representative of an action generated by at least one language model of one or more language models and a score representing a degree to which an outcome corresponding to execution of the action satisfied a corresponding objective;
• identifying, by the one or more processors, an indication of an outcome resulting from execution of a first action corresponding to a first data record of the plurality of data records, the first action comprising a tool invocation or control command that caused a measurable change in a simulated or physical environment;
• generating, by the one or more processors using a reward model, a score instance for the first action based on the indication of the outcome, the score instance representing a degree to which the outcome satisfied the corresponding objective;
• updating, by the one or more processors, the first data record in the database with an updated score for the first action generated based on the score instance generated based on the indication of the outcome;
• generating, by the one or more processors using a language model of the one or more language models, a search query for at least one data record in the database, the search query generated according to an input query;
• selecting, by the one or more processors, the first data record of the plurality of data records based at least on the updated score of the first data record and a similarity between the search query and the first data record;
• and
• executing, by the one or more processors, the language model using the first data record to generate an output action corresponding to the input query.

A non-transitory computer-readable medium storing instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising:

• maintaining a database storing a plurality of data records accessible to a language model, each data record comprising data representative of an action generated by at least one language model of one or more language models and a score representing a degree to which an outcome corresponding to execution of the action satisfied a corresponding objective;
• identifying an indication of an outcome resulting from execution of a first action corresponding to a first data record of the plurality of data records, the first action comprising a tool invocation or control command that caused a measurable change in a simulated or physical environment;
• generating, using a reward model, a score instance for the first action based on the indication of the outcome, the score instance representing a degree to which the outcome satisfied the corresponding objective;
• updating the first data record in the database with an updated score for the first action generated based on the score instance generated based on the indication of the outcome;
• generating, using a language model of the one or more language models, a search query for at least one data record in the database, the search query generated according to an input query;
• selecting the first data record of the plurality of data records based at least on the updated score of the first data record and a similarity between the search query and the first data record;
• and
• executing the language model using the first data record to generate an output action corresponding to the input query.