Self-Evolving Memory for Prompt Optimization (2026)

Free

Memory-augmented APO that stores historical refinement insights and reuses them across iterations

FreeFree tier

Inputs: textOutputs: text

Type

Open Source

Visit Website GitHub (40 stars)

About Self-Evolving Memory for Prompt Optimization (2026)

MemAPO is a memory-driven framework for automatic prompt optimization that reconceptualizes prompt optimization as generalizable and self-evolving experience accumulation. It maintains a dual-memory mechanism that distills successful reasoning trajectories into reusable strategy templates while organizing incorrect generations into structured error patterns that capture recurrent failure modes. For a new prompt, MemAPO retrieves both relevant strategies and failure patterns to compose prompts that promote effective reasoning while discouraging known mistakes. Through iterative self-reflection and memory editing, MemAPO continuously updates its memory, enabling prompt optimization to improve over time rather than restarting from scratch for each task. Experiments on diverse benchmarks show that MemAPO consistently outperforms representative prompt optimization baselines while substantially reducing optimization cost.

Key Features

Dual-memory mechanism: reusable strategy templates and structured error patterns for failure modes

Self-reflection and memory editing for continuous improvement over time

Retrieval of relevant strategies and failure patterns for composing effective prompts

Generalizable experience accumulation across tasks and queries

Substantially reduces optimization cost compared to baselines

Pros & Cons

Pros

Consistently outperforms representative prompt optimization baselines
Substantially reduces optimization cost through reusable knowledge
Generalizes across diverse tasks and queries without starting from scratch
Accumulates and improves prompting knowledge over time via self-reflection

Cons

Limited evaluation to benchmarks; real-world deployment details not provided
Requires initial memory bootstrapping (potentially from prior optimization runs)
Complexity of dual-memory management may introduce overhead

Best For

Automatic prompt optimization for large language models on downstream tasksAdapting LLMs to handle heterogeneous queries with reusable prompting knowledgeEfficient prompt engineering by learning from past successes and failures

FAQ

What is MemAPO?

MemAPO is a memory-driven framework for automatic prompt optimization that maintains a dual-memory mechanism: reusable strategy templates from successful reasoning trajectories and structured error patterns from incorrect generations. It retrieves relevant strategies and failure patterns to compose prompts that promote effective reasoning while discouraging known mistakes, and continuously updates its memory through iterative self-reflection.

How does MemAPO differ from existing prompt optimization methods?

Existing methods typically search for a specific prompt specialized to a fixed task, limiting generalization. MemAPO reconceptualizes prompt optimization as generalizable and self-evolving experience accumulation, enabling reuse of prompting knowledge across tasks and queries without restarting from scratch.

What are the main benefits of MemAPO?

MemAPO consistently outperforms representative prompt optimization baselines on diverse benchmarks while substantially reducing optimization cost. It allows prompt optimization to improve over time rather than restarting from scratch for each task.