You have probably thought about this now thoroughly that I have but...

1. Don't we need several types of memory? Short term for day to day transient data - like package numbers? Task memory for repeating cron tasks? Skills memory for new repeatable skills i.e. saved context for a shorter way to repeat a command? Knowledge - when Caal researched something, shouldn't it retain this knowledge? Isn't this how openclaw becomes so useful?

2. What technologies should be used for memory? Redis? Vector db? MCP?

3. AI hallucinations are partly due to lack of knowledge, but also due to false knowledge. What do we need to ensure that the memory is correct and factual?

4. Forgetting - in the long run forgetting is probably as important as remembering otherwise Caal is likely to grind to a halt. Some memories are only needed for a predefined period. Some are personal and unique, some are learned off the internet and can be forgotten and then relearned. Some are accurate when they are remembered but the world changes around them and they become inaccurate. Each memory probably needs metadata so that Caal knows when to forget.

5. Context - when I want to remember something I don't want my brain to be flooded with every memory however irrelevant - but equally I don't want to be starved of memory because I can't give it a precise index id. So memory needs to be fuzzy contextual. And some context can be precise i.e. if I am asking about a delivery, you only want details of all recent packages not my entire history - but others might be more general (like "I asked you about something to do with travel last week" when the history includes taxis, flights, hotels, reviewing car models, reviewing holiday destinations etc. and you got it wrong because it was 2.5 weeks ago).

IMO memory is likely to be the difference between Siri and openclaw i.e. limited Vs limitless.

In other words, it's complicated. But there has to be a lot of research on this, so probably not necessary to reinvent the wheel.

@Sophist-UK, Great comment and you're touching on something I’ve been thinking about a lot.

You're right that memory has layers. Short-term is step one - this PR covers transient data like flight numbers, package tracking, things that are useful for a few days and then expire. TTL-based, simple, predictable.

Long-term memory is planned as well. Thinking graph-based (something like Graphiti) with embeddings for contextual retrieval. This is where "Corey prefers morning flights" or relationships between contacts and preferences would live. Your points about forgetting and fuzzy contextual search are spot on for that layer - metadata-driven expiry and hybrid search (semantic + keyword) are likely where that heads. Trick is to retrieve that information when necessary and inject it.

Where CAAL's approach differs from what you might be picturing is the role memory plays. In CAAL's architecture, the LLM is a router - it decides which tool to call and with what parameters, and then deterministic n8n workflows execute. Memory serves that routing. When you say "is my flight on time?" the model needs to know which flight so it can call the right tool with the right parameters. It's not accumulating capability or learning new skills - memory serves as data to get enough context to make better routing decisions.

Skills in CAAL are n8n workflows. They go through review (automated + human) before they're live. The model can build new workflows (we showed this in a previous video) but it has to be prompted to do so, and the method is through calling another workflow that uses a larger LLM to generate the workflow. That boundary is intentional - it's what lets an 8B model be reliable and secure. The model doesn't need to be smart enough to self-improve, it needs to be smart enough to route.

So to your five points: 1 and 2 - yes, layered memory with graph + embeddings is on the roadmap. 3 - agreed, and scoping what the model can do with memory (route, not execute) helps bound that risk. 4 - absolutely, TTL is built into this PR and long-term will need smarter expiry. 5 - contextual retrieval is key for the long-term layer.

Appreciate the thoughtful input. This is exactly the kind of discussion that helps shape the architecture.

Any experience with Graphiti or similar?

cmac