Skip to main content
LLM Memory and Context Engineering home pageLLM Memory and Context Engineering
Framework

ACE Framework

Collaboration among Generator, Reflector, and Curator roles

ACE borrows the agentic design from Dynamic Cheatsheet, decomposing context adaptation into three roles.

Three Roles at a Glance

RoleInputOutput
GeneratorQuery + current PlaybookReasoning trace (including strategies and errors)
ReflectorTrace + execution feedbackSpecific experiences (insights)
CuratorExperiencesDelta context entries

Workflow

New query → [Generator] → Reasoning trace
                              ↓
                        [Reflector] → Experiences (multi-round refinement possible)
                              ↓
                         [Curator] → Delta entries
                              ↓
             Merge into Playbook (non-LLM logic)

Responsibilities of Each Role

Generator

  • Faces new problems, produces complete reasoning traces
  • Explicitly annotates which bullets are useful and which are misleading
  • Feedback guides the Reflector

Reflector

  • Extracts specific experiences from traces (successful strategies, failure patterns)
  • Supports multi-round iteration (default max 5 rounds)
  • Separated from Curator to avoid mixing "evaluation + organization" in one model

Curator

  • Merges experiences into compact delta entries
  • Uses deterministic non-LLM logic for merging, supporting parallelism
  • Ensures existing knowledge is not erased

Playbook Composition

Context is stored as a structured bullet list, with each bullet containing:
FieldContent
idUnique identifier
helpful counterNumber of times marked useful
harmful counterNumber of times marked misleading
contentA small reusable unit of strategy, concept, or failure pattern
Bullets are similar to memory entries in Dynamic Cheatsheet and A-MEM but add helpful/harmful counters to support subsequent deduplication and filtering.

Why Separate the Roles

  • Generator focuses on reasoning; model attention is not diverted
  • Reflector focuses on induction; avoids losing focus from "generating while summarizing"
  • Curator focuses on merging; uses non-LLM logic to ensure determinism and parallelism
Ablation experiments show that both the Reflector and multi-round refinement independently yield significant gains (see Ablation Study).

Offline vs Online Modes

ACE supports both modes simultaneously:
ModeScenarioDescription
OfflineOptimizing system promptsIteratively improves Playbook using training set
OnlineTest-time memoryPredicts first, then updates per sample
The two modes can be combined: Offline warmup → Online adaptation for further improvement.

Key Implementation Settings

ItemValue
Shared model for three rolesDeepSeek-V3.1 in non-thinking mode
Batch size1 (one delta generated per sample)
Reflector iteration limit5
Offline epoch limit5
Next: Explore two key mechanisms: Delta Updates and Grow-and-Refine.
LLM Memory and Context Engineering home pageLLM Memory and Context Engineering
Overview
Framework
Benchmarks
Analysis