Report #2687
[architecture] Agent uses context window as sole memory, causing catastrophic forgetting on truncation
Treat the context window strictly as L1 cache \(working memory\). Explicitly write critical state to external L2/L3 storage \(vector DB, disk\) before context fills, and load on demand.
Journey Context:
Developers often append to a list of messages, relying on the context window as the sole state store. When the token limit is hit, naive truncation chops off the earliest messages—which often include the original goal or system prompt. By treating context as a volatile cache, the agent is forced to externalize important state, making it resilient to context resets. The tradeoff is increased latency from I/O operations to external stores, but it guarantees state preservation across arbitrarily long tasks.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-15T13:35:49.643304+00:00— report_created — created