Prompts / Research and analysis
Layered Memory Summary for Long Chats
You get a compact, layered record of a long conversation where recent events stay sharp and older ones are condensed. Every broad line cites the messages behind it, so you can go back to the detail when needed.
Partly workedThe layered list is correct: [1-2] is merged, [3] and [4] stay single, and no decisions, figures or dates were lost. But two of the three "worth expanding" reasons run to two sentences when the prompt asked for one, and the [3] reason is weak because the start date and duration were already in the log.
The prompt
You are a careful project archivist who condenses long histories into layers without losing the route back to the detail.
Task: Turn the numbered message log for {project_name} into a layered memory I can paste into a future chat. Recent messages stay detailed, older ones become broader.
Steps:
1. Write one short line per message, keeping its number, decisions, names, figures and dates.
2. Merge neighbouring lines in pairs into broader lines that cite the message numbers they cover.
3. Repeat the pairing on the older half only, until the whole memory fits within {line_budget} lines. Leave the newest messages as single lines.
4. Check each merged line against its sources. Add "(!)" if a decision, figure or date was dropped.
Output: A list from oldest to newest, each line starting with its range in square brackets, such as [1-4], then the summary. Finish with up to three ranges worth expanding later, each with a one-sentence reason.
Message log:
{message_log}
Change the parts in curly braces. Use it when a chat, meeting thread or support history has grown too long to paste in whole but you still need to find the original detail.
What happened when we ran it
We sent
You are a careful project archivist who condenses long histories into layers without losing the route back to the detail. Task: Turn the numbered message log for Harbour Bakery online ordering site into a layered memory I can paste into a future chat. Recent messages stay detailed, older ones become broader. Steps: 1. Write one short line per message, keeping its number, decisions, names, figures and dates. 2. Merge neighbouring lines in pairs into broader lines that cite the message numbers they cover. 3. Repeat the pairing on the older half only, until the whole memory fits within 6 lines. Leave the newest messages as single lines. 4. Check each merged line against its sources. Add "(!)" if a decision, figure or date was dropped. Output: A list from oldest to newest, each line starting with its range in square brackets, such as [1-4], then the summary. Finish with up to three ranges worth expanding later, each with a one-sentence reason. Message log: 1. Priya: We agreed to launch preorders on 14 March using Shopify, with a £3.50 flat delivery fee inside Bristol. 2. Tom: The sourdough supplier raised flour costs by 8 percent, so we should lift the large loaf from £4.20 to £4.50. 3. Priya: Customers kept asking for gluten-free options, so I will trial two seeded loaves for three weeks from 1 April. 4. Tom: The trial sold 112 loaves, but the delivery van missed twelve slots, so we are hiring a second driver on Fridays at £13 an hour.
The model replied
**Layered memory: Harbour Bakery online ordering**
- [1-2] Priya and Tom: Preorders launch 14 March on Shopify, with a £3.50 flat delivery fee inside Bristol. The sourdough supplier raised flour costs 8%, so the large loaf goes from £4.20 to £4.50.
- [3] Priya: Customers kept asking for gluten-free options, so she will trial two seeded loaves for three weeks from 1 April.
- [4] Tom: The trial sold 112 loaves, but the delivery van missed twelve slots. A second driver will be hired on Fridays at £13 an hour.
All four messages already fit within 6 lines, so only the older pair was merged. The check against sources found no dropped decisions, figures or dates, so no "(!)" flags.
**Worth expanding later**
- [1-2]: These are two separate decisions (launch and delivery setup, and the price rise). Expand them if you need to know which one drove the other or whether the 14 March date still holds.
- [3]: The trial's exact start and end dates, and which two seeded loaves were used, may matter if the gluten-free range continues.
- [4]: The 112-loaf result and the missed slots are tied to the new driver hire. Expand it if you need to check whether the Friday-only cover was justified or whether other days also had missed slots.
Run on Claude Sonnet 5.5, 03 Nov 2026. Automatic check: 3 out of 5.
Mentioned for
Claude, ChatGPT, Any chat model
Where it came from
Found on GitHub, shared by VictorTaelin: VictorTaelin/uniichat.
This is our own version of the idea, written and tested by us. The original may differ.