This lesson covers adding persistent memory to MCP-based AI agents using three tools, save_memory, read_memory, and search_memory, backed by a shared JSON file. It lets a Researcher and Writer agent pair store insights, recall past learnings, and build on previous work across sessions instead of starting from zero each time.
What you will be able to do
- Build three MCP memory tools: save_memory, read_memory, and search_memory
- Set up a shared JSON file (shared_memory.json) that stores timestamped memory entries both agents can read and append to
- Have agents check memory before starting a task and save new learnings after finishing it
- Share memory across multiple agents so a Researcher and a Writer learn from each other's notes
- Explain the difference between a context window and persistent memory when designing agent systems
Before you start
- A working multi-agent MCP system with a Researcher and Writer agent sharing an MCP server, built in the previous lesson
- Basic understanding of how MCP tools are defined and called by agents
- A place to store a shared file that both agents can read from and write to
Reference
| Tool/Element | Input | Purpose | When to use |
|---|---|---|---|
| save_memory | Memory content (text describing what was learned) | Saves content to shared_memory.json with a timestamp | After completing a task, to store new insights |
| read_memory | None | Returns all saved memories | Before starting a new task, to check past learnings |
| search_memory | Search term (e.g. "writing style") | Returns memories matching that term | When an agent needs focused context on one topic instead of everything |
| shared_memory.json | N/A | Single file storing all memory entries; both agents read from and append to it | Acts as the shared filing cabinet for the whole agent team |
| Memory record | timestamp, agent, content | JSON structure of each saved entry | Written automatically every time save_memory is called |
Common errors and fixes
| What goes wrong | The fix |
|---|---|
| Assuming a large context window means the AI remembers past sessions | The context window only covers the current conversation; build a separate memory layer for anything that needs to survive between sessions |
| Agents re-research topics and repeat past mistakes every session | Call read_memory before starting work so agents build on what was already learned |
| Agents finish a task without recording what they learned | Call save_memory after completing significant work so learnings compound over time |
| Reading the entire memory file overwhelms the agent with unrelated information | Use search_memory with a specific term to get focused results instead of reading everything |
Read the full walkthrough
The complete lesson, with screenshots and any downloads, is published on Substack as part of MCP Masterclass: Connect AI to Everything.
More in this section
Continue the course
Browse all lessons in the MCP Masterclass: Connect AI to Everything course, or subscribe to the GenAI Unplugged newsletter to get new lessons in your inbox.
