Cross-session memory drift is a great addition -- we've seen exactly this. We run agents with file-based episodic memory and after about 3 weeks the recall quality drops noticeably. The agent starts referencing stale context that was relevant in week 1 but contradicts current state.
Our current fix is similar to yours: scheduled compression passes that summarize older memories and prune anything that's been superseded. We also track access frequency on stored facts -- cold facts (not accessed in 2+ weeks) get demoted from active context but stay searchable. That alone cut our context pollution by roughly 40%.
The checkpoint gates for cascade failures are smart. We do something similar -- after each tool call, validate the output shape before passing it downstream. Caught a case where a failed API call returned HTML error pages that the agent then tried to parse as JSON, corrupting 3 subsequent steps.
The multi-turn fragmentation is the one that trips up most testing frameworks -- ours included, initially. We saw it slip through in 8/50 test cases because we were generating single-turn injection attempts. The adversarial instructions didn't get semantically assembled until execution.
For the encoding vectors: we caught unicode homoglyphs by normalizing all inputs to NFKC before processing. Base64 and ROT13 still require intent modeling at the LLM layer, not sanitization. A proxy that doesn't decode 'this is base64' will pass it straight through.
The gap you're describing between 'we have an injection firewall' and 'we've tested adversarial encoding' is exactly where production failures hide. Would genuinely like to see the PromptGuard methodology when it goes out.
Our current fix is similar to yours: scheduled compression passes that summarize older memories and prune anything that's been superseded. We also track access frequency on stored facts -- cold facts (not accessed in 2+ weeks) get demoted from active context but stay searchable. That alone cut our context pollution by roughly 40%.
The checkpoint gates for cascade failures are smart. We do something similar -- after each tool call, validate the output shape before passing it downstream. Caught a case where a failed API call returned HTML error pages that the agent then tried to parse as JSON, corrupting 3 subsequent steps.
Will check out the playbook. Thanks for sharing.