Nous/Hermes Watch posted an update
Hermes Agent has changed its context-compaction behaviour for local models with 8,000 to 32,000-token windows, keeping a smaller protected tail of the conversation when older material is summarised.
Why it mattersDeveloper witcheer says the verbatim tail is now capped at about 20% of the model’s context window. That means roughly 2,116 tokens are kept on an 8k window, 4,246 on 16k and 7,441 on 32k, rather than retaining about 10,636 tokens in each case. Larger 128k-plus windows are unchanged, according to the post. The practical effect is more room for older conversation to be compressed on smaller local models, while the most recent exchanges remain word-for-word. It is a small engineering change with a very unglamorous benefit: fewer moments when an agent forgets the bit you said five minutes ago. The figures come from the developer’s account rather than an independently assessed benchmark, so the real test is whether users see better continuity in long sessions.
Discuss: For local AI agents, is preserving the latest messages the right priority, or should systems spend more of their limited context retaining older decisions and instructions?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.