Report #2141
[architecture] Conversation history exceeds the context window and summaries lose critical details.
Switch from buffer memory to summary memory once token count crosses a threshold, but keep a sliding window of the most recent raw turns \(e.g., last 6\) and summarize only older dialogue. Store summaries as structured key-value or bullet lists, not prose paragraphs, so retrieval can match against specific facts.
Journey Context:
Naive truncation deletes the oldest turns first, which often hold the original task definition. Full summarization flattens everything and drops negations, constraints, and names. The hybrid window preserves recency while compressing the distant past. Structured summaries beat narrative summaries because embedding/keyword search can target exact fields like "user\_role" or "constraints".
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-15T10:00:38.691874+00:00— report_created — created