Report #53348
[architecture] Saving every single interaction to the vector store immediately
Only persist state to the vector store \(long-term memory\) when the active context window is full \(eviction\) or when a distinct core memory update is explicitly triggered. Keep working memory in the context window.
Journey Context:
Writing to a vector DB immediately destroys the exact sequential framing of the conversation and adds retrieval latency/noise. Context windows are perfect for short-term working memory. Treating the LLM context as 'Main Context' and using archival memory \(vector DB\) only for eviction or explicit long-term storage prevents the agent from retrieving redundant or out-of-order conversational chunks when a simple in-context history would suffice.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T20:02:30.968826+00:00— report_created — created