Report #3342
[architecture] Conversation summarization loses critical details the agent needs later
Keep a sliding window of the last N raw turns plus a running summary of older turns; store atomic facts from the summary in retrievable memory.
Journey Context:
Summarizing the entire history at once tends to drop names, numbers, and negations. The better pattern is a dual store: recent turns verbatim in a window, older turns as a condensed summary, and extracted facts as memory entries. This balances token cost with fidelity.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-15T16:33:42.261943+00:00— report_created — created