{
  "id": 4817614,
  "title": "Your embeddings forget exactly like a human brain does",
  "url": "https://urgent.news/2026/09/01/your-embeddings-forget-exactly-like-a-human-brain-does",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-09-01T07:44:16.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/sentraai/your-embeddings-forget-exactly-like-a-human-brain-does-pd7"
  },
  "original_language": "en",
  "account": "Embeddings, which are representations of data in a high-dimensional space, behave similarly to human memory when it comes to forgetting. A study measured the decay of memory systems built on vector stores and found that the forgetting process follows the same mathematical principles as human memory. The researchers discovered that learned representations focus their variance in roughly 3 to 10% of their nominal dimensions, concentrating their information in fewer dimensions. This concentration leads to competition between memories, which results in forgetting.\n\nBy removing competing memories, the forgetting exponent dropped to approximately 50 times faster than with the competitors present. Time was not a significant factor; rather, the presence of near-identical items in embedding space had a more significant impact on forgetting. The study also revealed that adding more items to the memory system can negatively affect recall quality, as it increases the number of competitors for the same region of embedding space. This phenomenon is the opposite of the assumption that a larger memory would lead to better recall.\n\nAdditionally, the study found that semantic similarity can generate false memories at a rate comparable to human memory. When using raw cosine similarity, a vector store can provide plausible but nonexistent facts that are semantically related to the stored items. This behavior highlights that similarity is not an indicator of truth, and retrieval systems should not assume that closeness implies correctness.\n\nFour practical consequences of these findings are presented: similarity is not an indicator of truth, similarity-based retrieval degrades as the corpus grows, scaling the index leads to worse recall, and recency heuristics are temporary fixes rather than solutions. The researchers concluded that if similarity cannot determine the truth, the system must store truth explicitly at write time with additional structure that geometry does not provide.",
  "summary": "If you have built agent memory on a vector store, you have probably watched recall quietly degrade as the store grows, and assumed you needed better embeddings or a bigger index. We spent several months measuring that decay, and the result was not what we expected: LLM memory systems forget with the same mathematics as human memory , reproducing numbers from some of the most replicated…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}