LambdaDB Achieves 95% Input Token Savings in AI Memory Pipeline Using Native Embeddings

Details

LambdaDB team released an AI memory pipeline that preserves original text without knowledge graphs or LLM-based summarization, achieving 95% input token savings.
On LongMemEval (500 questions), the system achieved 96.0% accuracy (480/500) compared to 95.2% for full history, with average input tokens dropping from 105,183 to 4,886.
Estimated execution costs range from $7.51 to $9.42, a 91–93% reduction compared to the $105.53 cost of full-history processing.
The pipeline uses an ELT approach with three stages: indexing original text chunks with native embeddings, Bayesian hybrid search with native reranking, and context assembly within a 4,096-token budget.
BEAM benchmark results show scores of 89.00 at 100K tokens and 80.64 at 10M tokens, remaining within 3.1–4.4 points of past.dev public results.
Conversation branching and checkpointing leverage data versioning to reuse shared history, reducing memory preparation time to approximately 1.0 second for 500 messages versus 142 seconds for built-in vector storage.
SillyMemory is an open-source extension for SillyTavern that integrates LambdaDB storage, search, and versioning, requiring deactivation of built-in vector storage or summarization extensions.

Key point