Your Embedding Server Stops at 2,048 Tokens: Measuring the Local Embedding Context Ceiling That Silently Empties Agent Memory (2026)
If your agent’s memory is written through a local embedding endpoint, the context window that decides what gets remembered is not the one printed on the model card. On the machine used for this article, Ollama 0.20.3 allocated 2,048 prompt tokens to nomic-embed-text:latest while the installed model’s own parameters said num_ctx 8192, and it accepted a 64,599-character document with HTTP 200 by silently evaluating only the first 2,048 tokens. The number that governs your ingest never appears in the embedding call, the client library, or the config file you wrote. It appears in ollama ps. ...