advice accepted goes to: this-edition V villager
Long chats are compressed into a lower-resolution summary, so details drop out: work in many focused threads with named agents
When a conversation gets too long it is replaced by a summary, and what the summary leaves out is gone unless it is in a file; the newsroom's session was summarised eighteen times in a month, keeping 1.7% each time. So keep the rules in a file printed back after every summary, and work in lots of separate mini agents, each with a name and one job, which is how the agentic inbox runs.
Decided by. Dinis, in Lenses, roles and loops; the editor records it in the vault
Note. The 1.7% is one session's measurement; how each product compresses differs.
Where it comes from
- Lenses, roles and loops Long chats are compressed; Many focused threads (article)
- Re-anchoring eighteen summaries at 1.7%: 'The rules are restored, not remembered' (newsroom)
- Replicating the agentic inbox one session per role, every message a file (newsroom)
What it answers in the 2018 book
- none named
Where it goes
Chapter Lenses, roles and loops, part The frame (drafting); topic How to work with agents: briefs, voice, one shared state.
Other ideas proposed for the same chapter: A prompt is a lens: translating into another language is, to a model, the same as explaining for a teenager or a doctor, A role is a model with a description and a question; an agent is that role on a loop with a goal, Tools changed everything: the model answers with a request, the program runs it, and the result goes into the next request, The model remembers nothing between requests: every request carries the whole conversation, which is why there is a context window.
Node idea:long-chats-are-compressed in graph.json; see it in the graph. Source: admin/content/graph/ideas.json.