I've been testing the Claw beta for the last couple of weeks, mainly to see how its AI agents handle log correlation and anomaly scoring. The new 'memory' feature is interesting—it allows the analysis agent to persist context from one investigation case to another, supposedly to improve efficiency and pattern recognition.
My immediate low-level instinct is to ask: how is this memory isolated? If I'm working Case A (a sensitive internal fraud case) and then move to Case B (a routine phishing report), could contextual data from A leak into B's reasoning stream? Even if not full raw data, could inferred details influence the second case?
From a systems perspective, I'm thinking about:
* Is the memory a simple vector store per tenant, or is it scoped per analyst, per case group?
* What's the actual data flow? Is memory retrieval an explicit step the agent does, or is it automatic? If it's automatic, that's a bigger surface area.
* Could a cleverly crafted prompt in Case B trigger a "memory recall" that spills something from Case A?
```yaml
# Example from the beta config I'm testing
agent:
memory:
enabled: true
provider: "claw_vector_db" # proprietary
context_window: 10 # last N cases
```
The config suggests a rolling window, but "last N cases" is vague. Is that last N cases *I* worked on? Last N in the entire system? The docs aren't clear.
In a traditional SIEM, data is scoped by case or query. This feels more like a shared context buffer for the LLM agent. Could this become a weird side-channel? 🤔 I'm not even thinking about malicious intent yet—just accidental leakage because of how the underlying embeddings might work.
Has anyone else poked at this? I'm considering firing up some eBPF to trace the agent's calls to the memory backend during simultaneous case handling, just to see the scope of data fetched.
System calls per second matter.