Adaptive Recall: persistent memory system for AI assistants launches free tier

Adaptive Recall addresses a fundamental gap in current AI assistants: the inability to maintain and improve persistent context across conversations. While standard memory systems store text with vector embeddings and retrieve by cosine similarity, they are static and do not learn. Adaptive Recall applies three layers of sophistication—cognitive science (ACT-R scoring from 30 years of research), parallel search strategies, and automatic knowledge graph construction—to make memory retrieval progressively better.

The system's self-improving architecture is central to its design. Rather than relying on a single retrieval method, it runs four strategies in parallel and learns which to weight for different query types. Memories themselves evolve: they gain or lose confidence based on corroborating evidence and fade when no longer accessed, mirroring how human memory works. The ML pipeline trains on actual usage, ensuring that improvements are grounded in real patterns rather than generic assumptions. This approach contrasts sharply with static memory systems that treat stored information as unchanging rows in a database.

For developers, the barrier to entry is low: a free tier with 500 memories requires no credit card, and the API surface is simple (eight core tools). The choice of MCP as a primary transport (alongside HTTP REST) signals alignment with Claude and modern agent frameworks. The company's pending patent suggests confidence in the technical novelty of the approach. As AI assistants become more complex and multi-turn, persistent, learning memory may become a standard layer rather than an afterthought.