Skip to content

README.md

Latest commit

 

History

History
116 lines (90 loc) · 3.28 KB

README.md

File metadata and controls

116 lines (90 loc) · 3.28 KB

Memory Example

A memory system allows the agent to proactively store relevant information to be reused across different conversations.

Building Blocks

Instructions

We give the agent some additional hints on how to use the tools:

    const std::string instructions =
      "You are a helpful assistant with memory capabilities. "
      "You can remember information about the user using the write_memory tool"
      " and recall it later using the read_memory tool. "
      "When the user shares personal information (like their name, preferences,or important facts)"
      ", you must use write_memory to store it."
      "When needed, use list_memory to check if you have relevant stored memories."

Tools

This example implements a simple memory system (a single JSON file) with 3 tools:

  • list_memory: Lists all the keys currently stored.
  • read_memory: Given a key, reads a previously stored value.
  • write_memory: Writes information with a key-value pair.

Building

Important

Check the llama.cpp build documentation to find Cmake flags you might want to pass depending on your available hardware.

cd examples/memory

git -C ../.. submodule update --init --recursive

cmake -B build
cmake --build build -j$(nproc)

Using a custom llama.cpp

If you have llama.cpp already downloaded:

cmake -B build -DLLAMA_CPP_DIR=/path/to/your/llama.cpp
cmake --build build -j$(nproc)

Usage

./build/memory-example -m "path-to-model.gguf"

Example

Start one conversation and provide some personal information. The agent will use the write_memory tool to write to the memory:

$ ./build/memory-example -m ../../granite-4.0-micro-Q8_0.gguf
> My name is David and I love surfing
<tool_call>
{"name": "write_memory", "arguments": "{\n  \"key\": \"user_name\",\n  \"value\": \"David\"\n}"}
</tool_call>
[TOOL EXECUTION] Calling write_memory
[TOOL RESULT]
{"message":"Successfully stored memory with key 'user_name'","success":true}
<tool_call>
{"name": "write_memory", "arguments": {
  "key": "interest",
  "value": "surfing"
}}
</tool_call>
[TOOL EXECUTION] Calling write_memory
[TOOL RESULT]
{"message":"Successfully stored memory with key 'interest'","success":true}

If you close the previous one and start a new conversation, the agent can use the list_memories and read_memory tools to read the previously stored information.

$ ./build/memory-example -m ../../granite-4.0-micro-Q8_0.gguf
> What do you know about me?

<tool_call>
{"name": "list_memory", "arguments": "{}"}
</tool_call>
[TOOL EXECUTION] Calling list_memory
[TOOL RESULT]
{"keys":["interest","user_name"],"message":"Available memory keys:","success":true}
<tool_call>
{"name": "read_memory", "arguments": {"key": "user_name"}}
</tool_call>
[TOOL EXECUTION] Calling read_memory
[TOOL RESULT]
{"key":"user_name","success":true,"value":"David"}
<tool_call>
{"name": "read_memory", "arguments": {"key":"interest"}}
</tool_call>
[TOOL EXECUTION] Calling read_memory
[TOOL RESULT]
{"key":"interest","success":true,"value":"surfing"}
I have stored two pieces of information about you:

- **User name:** David
- **Interest:** Surfing

Let me know if there's anything else you'd like to add or discuss!