Jump to content

Connect Leeroopedia MCP: Equip your AI agents to search best practices, build plans, verify code, diagnose failures, and look up hyperparameter defaults.

Principle:CrewAIInc CrewAI Iterative Improvement

From Leeroopedia

Overview

A continuous improvement pattern where agent performance accumulates over time through memory subsystems that store execution history, entity relationships, and long-term learnings.

Description

Iterative Improvement uses the memory subsystem to help agents learn from past executions. Rather than treating each crew execution as an isolated event, the memory system creates a feedback loop where each execution informs future ones. This enables agents to progressively improve their outputs without changes to the underlying LLM or agent prompts.

The memory subsystem consists of three complementary stores:

  • ShortTermMemory — Stores recent interactions for the current session. This provides agents with immediate context about what has happened in the current execution, enabling them to reference earlier task outputs and avoid redundant work. ShortTermMemory uses vector-based storage (ChromaDB) for semantic similarity search.
  • LongTermMemory — Persists task-level insights across sessions. For example, "Agent X performed well on research tasks when given structured output requirements" or "Summarization tasks produce better results when preceded by a detailed research phase." LongTermMemory uses SQLite for durable, searchable storage of these learnings.
  • EntityMemory — Tracks relationships and facts about specific entities mentioned across executions. If the crew processes information about a company, person, or technology, EntityMemory stores structured knowledge about that entity that can be retrieved in future executions. EntityMemory uses vector-based storage for semantic retrieval.

Together, these three memory types create a layered learning system:

Memory Type Scope Storage Purpose
ShortTermMemory Current session ChromaDB (vector) Immediate context and interaction history
LongTermMemory Cross-session SQLite (relational) Task-level insights and performance patterns
EntityMemory Cross-session ChromaDB (vector) Entity relationships and factual knowledge

Theoretical Basis

This principle draws from two foundational concepts:

Experience Replay

Originally from reinforcement learning, experience replay stores past experiences (state, action, reward tuples) and replays them during training to improve learning efficiency. In CrewAI's adaptation, the "experiences" are task execution outcomes, human feedback, and entity knowledge. Instead of replaying for gradient updates, these experiences are retrieved via semantic search and injected into agent prompts as contextual information.

Continual Learning

Continual learning (also called lifelong learning) addresses how systems can learn from a continuous stream of data without catastrophically forgetting previous knowledge. CrewAI's memory subsystem implements a form of continual learning at the prompt level: new information is stored and retrieved alongside existing knowledge, allowing agents to accumulate capabilities over time without overwriting what was previously learned.

Improvement Mechanisms

The memory subsystem enables several improvement mechanisms:

  • Context enrichment — When an agent begins a task, relevant memories are retrieved and injected into the prompt, providing additional context that improves output quality
  • Pattern recognition — LongTermMemory stores task performance patterns that help agents learn which approaches work best for different task types
  • Entity consistency — EntityMemory ensures that agents maintain consistent knowledge about entities across executions, reducing contradictions and factual errors
  • Feedback integration — Human feedback from Training Execution is stored in memory, influencing future agent behavior

Relationship to Workflow

Iterative Improvement is the long-term value accumulator in the Crew Training and Testing workflow. It builds on Baseline Crew Configuration (which enables the memory subsystem), incorporates feedback from Training Execution, and its effects are measured through Performance Testing.

Implementation

Implementation:CrewAIInc_CrewAI_Memory_Subsystem

References

Page Connections

Double-click a node to navigate. Hold to expand connections.
Principle
Implementation
Heuristic
Environment