Journal
Blog
Research and product notes while we ship. Honest numbers. No vapor.
RSS
ThinkingRoot is the New State of the Art in LongMemEval Agent Memory Retrieval
ThinkingRoot achieves 99.0% Recall@15 on LongMemEval-S - the highest published retrieval accuracy on the benchmark - with only 481 tokens context and 16ms warm latency. Full methodology, graph architecture, and head-to-head comparison against Supermemory and Zep.
RotBench: the benchmark that catches AI memory rotting
Every memory product degrades as it accumulates - and nobody measures it. We built the first degradation-over-time benchmark, ran it on our own engine, and published the times it caught us being wrong.
CompAG: Compile-Augmented Generation
Every production memory system bets that understanding happens at query time. That bet has a ceiling. Here is why compile-augmented generation is the step after RAG.
In the pipeline
Coming soon- Engineering·Jun 24, 2026
Introducing Flows in ThinkingRoot
Coordinate multi-agent crews with persistent branching memory - fan out, compile, resolve drift.
- Product·Jun 18, 2026
Connected Memory: OAuth & API Integrations
Expose databases, files, and tools as secure agent functions - scoped per user session.
- Research·Jun 10, 2026
Building a Zero-Leak Accuracy Benchmark
Why traditional RAG benchmarks are easily prompted - and how a no-leak eval verifies real recall.