Unverified50% confidenceOpinionExact time
Randomly splitting the dataset by window causes adjacent windows to be highly similar, and the model may cheat by memorizing recording features, making the metrics artificially high.
1
Sources
50%
Confidence
Long-term
Relevance
11/13/2026
First Seen
Sources
Related Entities
Related Claims
VerifiedIf the memory store is full of duplicate, outdated, or mutually contradictory entries, the retrieval results will contain large amounts of low-value information, crowding out context space and even misleading the model, which is context pollution.75% similarUnverifiedHindsight-type memory components typically rely on vector databases, converting events into high-dimensional vectors via embedding models and retrieving them using cosine similarity or approximate nearest-neighbor algorithms.72% similarVerifiedThe larger the context window, the more information the model can remember in a single call, but it also brings rising inference costs and the problem of intermediate information being ignored.70% similarUnverified基准过拟合和数据污染会导致模型通过记忆而非真正理解获得高分,使公开基准分数可信度存疑68% similarUnverifiedHybrid approaches combining multiple memory sources performed worse than using a fact store alone, possibly due to signal conflicts among multiple memory sources.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1166350API
curl https://kongchang.com/api/v1/knowledge/claims/1166350MCP
get_claim(id=1166350)