The Heart of Context Engineering

August 8, 2025

WARNING: This post is over a year old. Some of the information this contains may be outdated.

🧠 AI Agent’s Memory: The Heart of Context Engineering

If you're building AI Agents, Memory is one of the most critical components. It enables agents to plan, act, and respond effectively — not just in the moment, but with awareness of past interactions and relevant knowledge.

Let’s break down the core types of memory in Agentic Systems:

🔁 1. Episodic Memory

This is where the agent stores past interactions — both user messages and its own responses/actions. These are usually embedded and stored in a Vector Database to support semantic search and recall.

🧩 Example: Saving chat history or executed steps in a task automation flow.

📚 2. Semantic Memory

This represents the agent’s knowledge base — internal docs, Notion pages, PDFs, or contextual grounding needed to isolate relevant information from large corpora.

💡 Think of it like the memory behind RAG (Retrieval-Augmented Generation).

⚙️ 3. Procedural Memory

This holds the system-level information: the structure of the system prompt, tool registry, and constraints/guardrails. It’s usually versioned and managed in Git or registries.

🔧 This is where your agent knows “how to think and act.”

💡 4. Pulling Memory into Context

When the agent needs to solve a task, relevant memory from Episodic, Semantic, or Procedural memory is retrieved and added to the context window.

🧠 5. Short-term (Working) Memory

All retrieved information is compiled into the final prompt passed to the LLM:

• Prompt structure
• Available tools
• Reasoning & action history
• Additional context

This is where real-time decision-making happens.

Comments are not enabled for this post.