Prompt optimization: learning from feedback

Memory can also change how an agent behaves. create_prompt_optimizer reads conversations and feedback and returns an improved system prompt.

Example

optimizer = create_prompt_optimizer(MemoryModel(), kind="prompt_memory")
trajectory = [
    {"role": "user", "content": "Please answer me by email."},
    {"role": "assistant", "content": "Here is your answer in the chat."},
]
new_prompt = optimizer.invoke({
    "prompt": "You answer support tickets for an online shop.",
    "trajectories": [(trajectory, "Use the contact method the customer asks for.")],
})
print(new_prompt)

A trajectory is a conversation, paired with feedback on it. The optimizer shows the model the current prompt, every trajectory and its feedback, and asks for the full updated prompt. The stand-in's improve_prompt only appends the feedback as a rule; a real model rewrites the prompt so the rule fits in.

Three kinds

kind	Model calls	How it works
`prompt_memory`	1	One request: current prompt, trajectories and feedback in, new prompt out.
`metaprompt`	Up to `max_reflection_steps` + 1	Reflects on the trajectories, possibly several times, before writing the prompt.
`gradient`	2 or more	First writes a critique of what went wrong, then applies it as edits to the prompt.

Only prompt_memory is run here: the other two ask the model for open-ended critique and reflection, which only a real model produces usefully. Their code is identical apart from kind.

Review before you ship

An optimized prompt is a new version of your agent. Keep it in version control, and run your evals, as in the DeepEval course, before switching to it.

Try it yourself

Pass two trajectories with different feedback.
Pass feedback of an empty string and compare the result.
Use create_multi_prompt_optimizer in LangMem's reference for an agent with several prompts.

PreviousSummarizing long conversations

Little by little, you're building something great.