Conversation Memory
Chat is the default UI. If history is messy, the agent contradicts itself mid-thread.
A meeting transcript with names on each line — not a pile of unmarked quotes.
Visual Workflows
Start here — scroll inside each diagram frame to explore, then use + / − to zoom up to 200% if needed.
Overview
Scroll inside the frame to explore · use + / − to zoom up to 200%
Scroll inside the frame to explore · use + / − to zoom up to 200%
Pack the next call
Scroll inside the frame to explore · use + / − to zoom up to 200%
Pin the goal, keep recent turns, summarize the tail.
Key Takeaways
- 1.Conversation memory is the chat thread: roles, turns, and order. Strategies: keep last N, summarize older, or pin key turns.
- 2.Role tags matter — user vs assistant vs tool must stay intact. A broken thread makes the model invent a past that never happened.
- 3.Store turns with role and id. Trim from the middle or summarize the oldest block.
- 4.Never merge user and tool messages into one blob.
Learn elsewhere
- →Working Memory
- →Memory Summarization
Real Example
Scenario
A 60-turn design chat: pin the original brief, keep the last eight turns, replace turns 1–50 with a one-paragraph summary.
What you would do
In Agent Memory, apply Conversation Memory to this scenario: A 60-turn design chat: pin the original brief, keep the last eight turns, replace turns 1–50 with a one-paragraph summary. Identify the inputs, run the technique, validate the output, and note one thing you would monitor in production.
Commands
Commands to Remember
Keep roles on every turnLast N + summaryPin the goalDo not flatten tools into user text
Cheat Sheet
Quick recap
quick ref- •Thread with names
- •Last N + summary
- •Pin what matters
- •Roles stay intact
Common Mistakes
- ✕Dropping tool messages from history
- ✕Summarizing so hard the user's constraint disappears
- ✕Sending the entire 200-turn log every time