Show HN: Recalld – A memory layer for AI agents that returns only relevant facts
a day ago
- Recalld is a memory layer for AI agents that extracts, updates, and retrieves facts from conversations and documents, eliminating the need for custom vector databases or retrieval pipelines.
- The API consists of three primitives: 'add' (ingest and reconcile facts), 'recall' (curated retrieval with LLM selection), and 'search' (raw dense-vector lookup).
- Recalld's curated recall returns only the facts that answer a query, using ~240 tokens on average, and achieves 88.7% accuracy on the LoCoMo benchmark with 6.7× fewer tokens than raw search.
- Key features include time-aware facts, automatic fact updates/supersession, bring-your-own-key support, active memory windows, and regional data residency (EU or US).
- Recalld integrates seamlessly as an MCP server, allowing any MCP-compatible agent to use its memory tools without SDK or glue code.
- GDPR-compliant by design: data stays in-region, supports right to erasure, encryption (TLS 1.3, AES-256), and provides a DPA and transparent sub-processor list.
- Pricing is credit-based with a free tier, Pro, Scale, and Enterprise options; search is ~40× cheaper than curated recall.