Adaptive Recall

Adaptive Recall is a memory layer for AI applications and agents. It stores what a model picks up across sessions, brings the right pieces back when they're relevant, and quietly improves the quality of that retrieval the more it gets used. Rather than treating memory as a static bucket of embeddings you query once and forget, it treats recall as something that should adapt to how information actually gets used over time. The problem it goes after is the ceiling most memory APIs run into. A lot of them do exactly one thing, cosine similarity over vectors, and hand back whatever sits closest in embedding space. That works for a first pass, but it misses recent context, exact keywords, and the relationships between facts, so agents end up with context that's technically similar and practically useless.

Under the hood it runs four retrieval strategies in parallel and blends the results into one ranked set. There's vector similarity for meaning, temporal recency for what happened lately, full-text keyword search for exact terms, and knowledge graph traversal for facts that connect to each other. Sitting on top of all four is a cognitive scoring model based on ACT-R, a line of psychological research into how human memory works that's been refined over roughly thirty years. The goal is to rank stored memories closer to the way human attention and recall behave, so items that keep proving useful rise to the top while stale ones fade into the background. No single signal gets to decide on its own, which is the core of the design.

It also builds a knowledge graph on its own, extracting entities and the relationships between them from whatever you store, so recall can follow connections instead of returning isolated snippets. Memories have a lifecycle here. Confidence in a given fact can grow or decay as time passes, and a consolidation process runs on a schedule to reorganize and clean up what's held. There's also a feedback tool that lets the system learn whether a particular recall actually helped, and that signal feeds a machine-learning pipeline that tunes ranking against real usage patterns rather than a fixed formula. Self-monitoring quality checks watch the whole thing, so the store is meant to get more accurate the longer an application runs against it instead of slowly filling with noise.

Developers reach it two ways. There's an MCP server so Claude Code and other MCP-aware command-line tools can call it directly, and there's an HTTP REST API secured with bearer tokens for everything else. The surface is small on purpose, eight core tools that cover store, recall, update, forget, graph, status, snapshot, and feedback. That compact set keeps it easy to wire into an agent without learning a sprawling SDK, and the same eight calls behave the same way whether you're driving them from an editor or from your own backend. A Memory Explorer gives you a view into what's been saved, and higher tiers add a Graph Explorer and a Data Browser for inspecting the entities and raw records the system is working with.

It's aimed squarely at people building AI agents and assistants that need to remember things reliably between runs, whether that's a coding agent, a support bot, or a personal assistant that should recall what you told it last week. If you've pointed a plain vector database at an agent and watched it surface context that matched on wording but missed the point, this is the gap it's trying to close. What sets it apart is the refusal to lean on one score. Recency, exact keywords, semantic meaning, and graph structure all get a vote, and a cognitive model informed by decades of memory research arbitrates between them. That combination tends to produce recall that feels less brittle than similarity search alone, especially as the amount of stored history grows.

Access is freemium. The free tier gives you 500 memories, twenty requests a minute, and consolidation every twelve hours, with no credit card required to get started. Paid plans lift those ceilings as you grow. Starter is $19.99 a month for 5,000 memories, sixty requests a minute, and consolidation every three hours. Pro is $49.99 a month for 25,000 memories, two hundred requests a minute, hourly consolidation, the Graph Explorer, and human support. Business runs $99.99 a month for 100,000 memories, six hundred requests a minute, and a fifteen-minute consolidation window. MCP and REST access are included on every plan, so the way you integrate stays identical from the free tier through production, and only the limits and support level change as you move up.