Mojentic.Agents.BaseLLMAgentWithMemory (Mojentic v1.2.0)

An LLM agent that uses SharedWorkingMemory to remember information.

This agent extends BaseLLMAgent with memory capabilities, allowing it to:

Access shared working memory in its system context
Learn new information from conversations
Merge learned information back into shared memory

The agent automatically includes memory context in its prompts and extracts memory updates from LLM responses using structured output.

Usage

alias Mojentic.Context.SharedWorkingMemory
alias Mojentic.Agents.BaseLLMAgentWithMemory
alias Mojentic.LLM.Broker

# Create shared memory
memory = SharedWorkingMemory.new(%{
  "User" => %{"name" => "Alice"}
})

# Create broker
broker = Broker.new("qwen2.5:7b", Ollama)

# Create agent with memory
agent = BaseLLMAgentWithMemory.new(
  broker: broker,
  memory: memory,
  behaviour: "You are a helpful assistant.",
  instructions: "Answer questions and remember new facts.",
  response_model: %{
    "type" => "object",
    "required" => ["text"],
    "properties" => %{
      "text" => %{"type" => "string", "description" => "Your response"}
    }
  }
)

# Generate response - memory is automatically included
{:ok, response, updated_memory} =
  BaseLLMAgentWithMemory.generate_response_with_memory(agent, "What's my name?")

Examples

# Create agent that remembers user preferences
memory = SharedWorkingMemory.new(%{})
broker = Broker.new("qwen2.5:7b", Ollama)

agent = BaseLLMAgentWithMemory.new(
  broker: broker,
  memory: memory,
  behaviour: "You are a personal assistant.",
  instructions: "Help the user and remember their preferences.",
  response_model: %{
    "type" => "object",
    "required" => ["response"],
    "properties" => %{
      "response" => %{"type" => "string"}
    }
  }
)

{:ok, response, memory} =
  BaseLLMAgentWithMemory.generate_response_with_memory(
    agent,
    "I prefer dark mode and vim keybindings"
  )

# Memory now contains user preferences

Summary

Types

t()

Functions

create_initial_messages(agent)

Creates initial messages with behaviour, memory, and instructions.

generate_response_with_memory(agent, content)

Generates a response with memory context and updates.

new(opts)

Creates a new BaseLLMAgentWithMemory.

update_memory(agent, memory)

Updates an agent's memory reference.

Types

t()

@type t() :: %Mojentic.Agents.BaseLLMAgentWithMemory{
  behaviour: String.t(),
  broker: Mojentic.LLM.Broker.t(),
  instructions: String.t(),
  memory: Mojentic.Context.SharedWorkingMemory.t(),
  response_model: map(),
  tools: [module()] | nil
}

Functions

create_initial_messages(agent)

Creates initial messages with behaviour, memory, and instructions.

The messages include:

System message with agent behaviour
Memory context (what the agent remembers)
User message with instructions

Parameters

agent - The BaseLLMAgentWithMemory instance

Returns

List of Message structs.

Examples

messages = BaseLLMAgentWithMemory.create_initial_messages(agent)
#=> [
#     %Message{role: :system, content: "You are..."},
#     %Message{role: :user, content: "This is what you remember:..."},
#     %Message{role: :user, content: "Answer questions..."}
#   ]

generate_response_with_memory(agent, content)

Generates a response with memory context and updates.

This function:

Creates initial messages with current memory
Adds the user's content
Requests structured output with memory field
Merges memory updates back into SharedWorkingMemory
Returns the response and updated memory

Parameters

agent - The BaseLLMAgentWithMemory instance
content - The user's input string

Returns

{:ok, response_data, updated_memory} on success
{:error, reason} on failure

Examples

{:ok, response, memory} =
  BaseLLMAgentWithMemory.generate_response_with_memory(
    agent,
    "My favorite color is blue"
  )

# response contains the structured response (without memory field)
# memory is updated SharedWorkingMemory with new information

new(opts)

Creates a new BaseLLMAgentWithMemory.

Parameters

opts: Keyword list with:
- :broker - LLM broker instance (required)
- :memory - SharedWorkingMemory instance (required)
- :behaviour - System message defining agent personality (required)
- :instructions - Instructions for processing events (required)
- :response_model - JSON schema for structured output (required)
- :tools - List of tool modules (optional)

Examples

agent = BaseLLMAgentWithMemory.new(
  broker: broker,
  memory: memory,
  behaviour: "You are a helpful assistant.",
  instructions: "Answer questions and learn new facts.",
  response_model: %{
    "type" => "object",
    "required" => ["answer"],
    "properties" => %{
      "answer" => %{"type" => "string"}
    }
  }
)

update_memory(agent, memory)

Updates an agent's memory reference.

Returns a new agent struct with the updated memory.

Parameters

agent - The BaseLLMAgentWithMemory instance
memory - The new SharedWorkingMemory instance

Examples

agent = BaseLLMAgentWithMemory.update_memory(agent, new_memory)