AgentSkillsCN

cognitive-architectures

为LLM智能体设计、优化并管理认知架构，从历史渊源到现代CoALA实现。 在以下场景中使用：构建自主智能体、设计记忆与规划系统、研究智能体行为，或优化智能体性能。

SKILL.md
--- frontmatter
name: cognitive-architectures
description: |
  Design, optimize, and manage cognitive architectures for LLM agents, from historical roots to modern CoALA implementations.
  Use when: building autonomous agents, designing memory/planning systems, researching agentic behaviors, or optimizing agent performance.

Cognitive Architectures

Cognitive architectures provide the structural design for AI agents, moving beyond simple "prompt-response" loops to stateful, goal-oriented entities.

How to Apply CoALA Framework

The CoALA (Cognitive Architectures for Language Agents) framework organizes agent design into four modules:

  1. •Memory:
    • •Working Memory: The active context window (current interaction).
    • •Episodic Memory: Past experiences (Vector DB/RAG).
    • •Semantic Memory: Facts and knowledge (Knowledge Graphs/Docs).
    • •Procedural Memory: Code/Tools/Skills (Implicit or Explicit).
  2. •Action Space: The set of executable external actions (API calls, tool use) and internal actions (memory retrieval, reasoning).
  3. •Decision Making: The "loop" (e.g., ReAct, OODA, Plan-and-Solve) that selects actions based on memory and perception.
  4. •Perception: How the agent inputs information (Text, Vision, Audio).

How to Optimize Performance

  • •Memory Retrieval: Use Hybrid Search (Keyword + Vector) to reduce retrieval latency and increase relevance. Implement Memory Decay to prioritize recent or frequently accessed memories.
  • •Reflection: Periodically summarize episodic memory into high-level insights (Semantic Memory) to prevent context saturation.
  • •Tool Use: Minimize token usage by optimizing tool descriptions and using structured outputs (JSON) for tool arguments.

Common Pitfalls & Solutions

PitfallImpactSolution
Context OverflowModel hallucination or truncation.Implement summarization hierarchies and strict token budgeting.
Action LoopsAgent gets stuck repeating the same tool call.Add a "max_retries" counter and a "give_up" or "ask_human" fallback.
Memory PollutionIrrelevant info clogs retrieval.Use strict filtering/ranking during RAG retrieval; periodically prune vector DB.

Evolution & History

Understanding the lineage helps choose the right pattern:

See History of Cognitive Architectures for the evolution from Symbolic AI to LLMs.

Examples

Example: Implementing a Reflection Step

Input: 10 recent user interactions (Episodic Memory). Goal: Update User Profile (Semantic Memory).

Prompt Pattern: "Review the last 10 interactions. Extract 3 key facts about the user's preferences. Update the 'User Profile' JSON object. discard the raw logs."

References