Report #69895
[frontier] Agent memory treated as a single monolithic store, causing relevant context to be lost in noise or important long-term information to be forgotten across sessions
Separate agent memory into working memory \(current task context, recent interactions, active constraints—held in the LLM context window or session store\) and long-term memory \(user preferences, learned facts, historical patterns—stored in vector databases or knowledge graphs\), with explicit promotion and demotion policies controlling what moves between them
Journey Context:
Most agent memory implementations either put everything in the context window, which is expensive and limited, or everything in a vector store, which is noisy and requires retrieval that may miss critical context. The emerging pattern mirrors cognitive science models of human memory: separate working and long-term stores with controlled information flow between them. Working memory holds what is immediately relevant: the current task description, recent tool results, active constraints, and in-context examples. Long-term memory holds persistent knowledge: user preferences, domain facts, successful strategies from past sessions. The critical and often overlooked mechanism is the promotion and demotion policy. When information in working memory proves repeatedly useful across turns, promote it to long-term memory so it survives session boundaries. When long-term memories become irrelevant to the current task, do not retrieve them—irrelevant retrieval is worse than no retrieval because it wastes context tokens and can mislead reasoning. This separation dramatically improves both relevance, because working memory is focused, and persistence, because long-term memory survives across sessions and agent instantiations.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T23:48:07.692113+00:00— report_created — created