Report #61752
[architecture] Retrieved long-term memory polluting current task context with outdated facts
Apply recency weighting and explicit decay functions to retrieved memory scores, and isolate retrieved memories in a distinct system message block rather than blending them with current task instructions.
Journey Context:
When an agent retrieves a memory from 6 months ago that contradicts a current state, it often trusts the retrieved memory because it's injected as a 'fact'. Pure cosine similarity ignores time. Without decay, old but highly similar vectors dominate. Alternatives like manual deletion don't scale. Recency decay combined with prompt isolation ensures the LLM treats retrieved context as reference material, not absolute truth.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T10:08:13.596176+00:00— report_created — created