Short Term Memory is Live! 🚀

[cmac86]

CAAL now remembers context between tool calls.

When you ask "where is my package?" CAAL can pull your package tracking number from memory and chain it into the package_tracker tool - no need to repeat yourself. And when a tool returns useful data, it can store it automatically for later use.

What's New

Three Ways to Store Memories

1. Tool hints - n8n workflows can return a memory_hint in their response. CAAL stores it automatically. Your flight tracker returns UA1234, and next time you ask about your flight, CAAL already knows.

2. Explicit storage - Tell CAAL to remember something. "Remember my flight is UA1234" stores it via the memory_short tool. Store, get, delete, list - all through voice.

3. HTTP API - POST /memory for external systems. Your n8n workflows or other services can push context into CAAL's memory.

Memory Panel

New Brain icon in the frontend. View all stored memories with source badges (tool/voice/api), timestamps, and TTL countdown. Inline edit and clear all.

TTL-Based Expiry

Memories expire automatically. Flight numbers, package tracking, delivery updates - useful for a few days, then gone. No manual cleanup needed. Editable in the UI and a parameter in memory_short tool so you can tell CAAL how long to remember it for.

How It Works

Currently, short-term memory is injected into context so the model is aware of what's stored. The model also has a memory_short tool to explicitly fetch, store, and manage memories. Once caal-ministral is trained on memory tool calls, the auto-injection will be removed to reduce context bloat. The model will learn to fetch from memory when it needs to, rather than having everything injected on every call.

This is short-term memory - transient data that helps the model make better routing decisions. It's not learning new skills or accumulating intelligence. It gets enough context to call the right tool with the right parameters.

What's Next

Thinking about long-term memory as the next layer. Some early ideas we're exploring:

• Tool-scoped memory - tag memories with the tools they relate to, so the model only gets relevant context when it needs it
• Pattern detection - analyze tool call history to surface defaults ("I notice you always email John Smith when you say 'email John' - want me to remember that?")
• Graph-based and vector based storage - persistent knowledge for preferences, contacts, relationships

If you have thoughts on how memory should work, share them in this thread.

Try It

Pull the latest and start telling CAAL things to remember.

Video

A deeper walkthrough video is in the works covering the full memory system and the architecture thinking behind it.

[nerd8east-a11y]

Hello, Do you mind showing me the proper commands to get the latest version of CAAL?
Last time I updated was with these lines from our previous dicussion in here:

cd CAAL
git fetch origin
git checkout feat/tools-registry
docker compose --profile https up -d --build
or
docker compose up -d --build

Not sure if this is still the best way to update my code.
Sorry I am still learning about github and everything else.
Thanks

[AbdulShahzeb]

@nerd8east-a11y

cd CAAL
git checkout main
git pull
docker compose up -d

HTTPS is now default.

[nerd8east-a11y]

thanks. I will give this a try.