{"slug": "your-embeddings-forget-exactly-like-a-human-brain-does", "title": "Your embeddings forget exactly like a human brain does", "summary": "A developer's measurements show that LLM memory systems forget with the same mathematics as human memory, reproducing classic psychological results without tuning. The team found that embeddings have an effective dimensionality of around 16, causing memories to compete and interfere, and that false-memory rates on raw cosine similarity match human rates. They conclude that vector stores cannot distinguish true from false, so systems must record truth explicitly at write time.", "body_md": "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.\n\nWe 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 experiments in clinical psychology. No tuning required to get there.\n\nStart with the finding everything else follows from. Take an embedding model that advertises 384 or 1,024 dimensions and measure where the variance actually lives:\n\n``` python\nimport numpy as np\n\n# X: (n_samples, n_dims) matrix of embeddings from any pretrained model\nX = X - X.mean(axis=0)\neigenvalues = np.linalg.svd(X, compute_uv=False) ** 2\np = eigenvalues / eigenvalues.sum()\n\n# participation ratio: how many dimensions are doing real work\neffective_dims = 1.0 / np.sum(p ** 2)\nprint(f\"nominal: {X.shape[1]}, effective: {effective_dims:.1f}\")\n```\n\nRun that on a model advertising 384 to 1,024 dimensions and you get an effective dimensionality around **16**. Learned representations concentrate their variance into roughly 3 to 10% of their nominal dimensions.\n\nThat is not a defect of one model. It is a property of learned representations, and it is why compression works at all. It is also why memory built on those representations behaves like a crowded room rather than a filing cabinet: with 16 effective dimensions, every new memory lands close to existing ones, and closeness is interference.\n\nThe Ebbinghaus forgetting curve is usually taught as memory fading over time. Our measurements say the mechanism is different: **memories compete, and competition looks like decay.**\n\nThe test is simple. Measure the forgetting exponent normally, then remove the competing memories and measure again:\n\n``` php\nwith competitors present     ->  power-law forgetting, exponent ~ human curve\ncompetitors removed          ->  exponent drops ~50x\n```\n\nFifty-fold. Time barely matters; neighbours matter enormously. Which means the practical lever on agent memory recall is not retention windows or TTLs, it is **reducing how many near-identical items compete for the same region of embedding space.**\n\nIf your retrieval quality fell off a cliff after you tripled the corpus, this is why. You did not lose information, you added competitors.\n\nThe part that genuinely surprised us. The classic false-memory experiment (a lure word that was never presented gets \"recalled\" because it is semantically central to the list) reproduces on raw cosine similarity over unmodified pre-trained embeddings:\n\n| Measurement | Rate |\n|---|---|\n| Human false-memory rate, classic studies | ~0.55 |\n| Raw cosine similarity, no tuning | 0.583 |\n\nZero parameter fitting. Nobody engineered this. Semantic similarity alone manufactures confident recollections of things that were never stored, at approximately the human rate.\n\nRead that back as an engineering statement: **a vector store will hand your agent a plausible fact that was never written down, and it will look exactly like a real retrieval.** No confidence score distinguishes them, because from the geometry's point of view there is nothing to distinguish.\n\nFour consequences, in the order they will bite you:\n\nThe architectural conclusion we drew, and the reason we build what we build: if similarity cannot tell you what is true, the system has to record truth explicitly, at write time, with structure that geometry does not provide.\n\n```\n{\n  \"statement\": \"Acme's latency fix slipped to Q3\",\n  \"valid_from\": \"2026-04-03\",\n  \"valid_to\": null,\n  \"supersedes\": \"fact_8812\",\n  \"source\": \"meeting:2026-04-03#turn-58\",\n  \"visible_to\": [\"role:account-team\"]\n}\n```\n\nThree fields there do work that no embedding can do. `valid_from`\n\nand `valid_to`\n\nmake time explicit rather than inferred. `supersedes`\n\nrecords that a previous belief was replaced, so the old one can be retired instead of competing forever. `source`\n\nmakes the claim checkable.\n\nNone of that is a better vector. It is a different data model, and it exists precisely because the geometry has the failure modes above.\n\nThe participation-ratio snippet above runs on any embedding matrix in about three lines. If your effective dimensionality comes back in the teens while you are paying for 1,024, you now know why your recall curve looks like a psychology textbook.\n\nFull methodology, the compression results behind the 3 to 10% figure, and the rest of the experiments are in [the original writeup](https://www.sentra.app/blog/geometry-of-forgetting). If you want the practical version, we wrote up [why embeddings alone are not memory](https://www.sentra.app/articles/embedding-models-and-ai-memory) and [what breaks when retrieval is treated as memory](https://www.sentra.app/articles/why-rag-fails).\n\n*This research came out of building Sentra, a company brain for teams and AI agents. We went looking for a compression result and found a psychology paper instead.*", "url": "https://wpnews.pro/news/your-embeddings-forget-exactly-like-a-human-brain-does", "canonical_source": "https://dev.to/sentraai/your-embeddings-forget-exactly-like-a-human-brain-does-pd7", "published_at": "2026-09-01 07:44:16+00:00", "updated_at": "2026-09-01 07:54:03.537550+00:00", "lang": "en", "topics": ["large-language-models", "ai-research", "ai-agents", "ai-infrastructure"], "entities": [], "alternates": {"html": "https://wpnews.pro/news/your-embeddings-forget-exactly-like-a-human-brain-does", "markdown": "https://wpnews.pro/news/your-embeddings-forget-exactly-like-a-human-brain-does.md", "text": "https://wpnews.pro/news/your-embeddings-forget-exactly-like-a-human-brain-does.txt", "jsonld": "https://wpnews.pro/news/your-embeddings-forget-exactly-like-a-human-brain-does.jsonld"}}