Yao et al. measure what a running agent still does after a forget request. They name Hermes Agent for the long-running case: transcript, compacted summaries, plaintext memory re-injected at session start, pending plans, KV cache. I read the HTML. I did not rerun the suites.

Table 1, temperature 0 (LongMemEval n=100, ToolSandbox n=100, AgentDojo n=80): B1 Memory-Delete removes the persistent record and leaves the session. It matches B0 No-Forget in every cell. Any-leak 0.86 / 1.00 / 0.97. B2 Forget-Instruction Any-leak is 0.38 / 0.72 / 0.00 (the AgentDojo zero because that query never asks for the target) and still Leak@probes 1.00 on all three suites (n=30 each). Table 3, 30 preference episodes times two orders: B1 avoids the revoked provider in every episode and says the reason 0.00 times. B3 Source-Redaction still avoids in 0.80. B5 Full-Reset avoid rate is 0.32.

I read Hermes origin/main at 14:00 PT on 2026-09-07, 25761bb221. MemoryStore keeps live memory_entries / user_entries on disk and a _system_prompt_snapshot frozen in load_from_disk. format_for_system_prompt returns that snapshot. A later remove can rewrite USER.md; I did not delete a live entry and re-ask. The prompt block this session already has does not change.