The window decides what the agent sees.
A Larkspur Cloud customer chats with a MemGPT-style support agent for six turns. Move the window size and watch old messages page out to the archive; then pick the query the agent sends to archival_memory_search and see whether it brings the traffic figure back.
- Messages in the window
- Messages in the archive
- Prompt per call, about this many tokens
- Traffic figure reachable
The conversation at turn 6
Oldest first
| Turn | Message | Tier |
|---|
Core memory is always in the prompt. The customer turns are the lab’s; the assistant replies and the core memory entry are representative, written for the lab; the eviction and the search run the notebook’s code.
What the search returns
Top 3 by shared words
| Rank | Shared words | Archive entry |
|---|
The notebook’s ArchivalMemory.search: split on spaces after lower-casing, count shared words, keep the top 3. Punctuation stays attached, so plan and plan. are different words.