Recalld launches a long-term memory layer for AI agents

Summary

Recalld is a memory service for AI agents that stores conversations and documents, extracts atomic facts, anchors them in time, and updates, supersedes or merges older facts when new information arrives. Its API centers on three operations: adding memory, curated recall and raw search. Curated recall combines hybrid retrieval with an LLM selection pass and returns a short ranked set of facts, while search returns dense-vector candidates for the customer’s own model to filter. On the company’s LoCoMo evaluation, recall achieved 88.7% answer accuracy versus 88.2% for search, while returning an average of 243 tokens instead of 1,627, or 6.7 times less context. The evaluation covered 1,540 non-adversarial questions across 10 conversations and reported 80.5% accuracy on multi-hop questions, with lower-level results of 91.7% for single-hop, 89.7% for temporal and 83.3% for open-domain questions. Recalld says the benchmark used an external Agent Memory Benchmark harness with unchanged prompts, but notes that its judge may accept “I don’t know” as correct, so the accuracy should be treated as an upper bound. The service is available in EU and US regions, offers GDPR-related controls, deletion, export, encryption and data-processing documentation, and can be connected through a native Model Context Protocol server. Plans range from a free tier to $99 per month, with credit-based billing, different active-memory windows and BYOK support on Pro and Scale plans. Recalld also offers a chat assistant built on the same engine.