RecallDB: A Local-First, Bitemporal Hybrid Engine for Long-Horizon Agent Memory and Decoupled Evaluation

Autonomous artificial intelligence (AI) agents operating over extended multi-month or multi-year horizons require persistent, verifiable episodic and semantic memory. Existing external agent memory frameworks rely almost exclusively on flat, atemporal vector databases, suffering from temporal blindness and contradiction collapse. In this paper, we introduce RecallDB, a zero-daemon, local-first embedded memory engine that formalizes agent state persistence through a bitemporal relational and lexical substrate. RecallDB couples an embedded SQLite engine operating in Write-Ahead Logging (WAL) mode with an FTS5 BM25 lexical inverted index and an IEEE 754 single-precision float32 vector BLOB index. Contradiction resolution is governed by a deterministic, atomic Supersession Finite State Machine (ACTIVE -> SUPERSEDED) that preserves an unbroken historical audit trail. Evaluated on the SynTemp-100 longitudinal benchmark, RecallDB achieves 1.0000 Recall@1 (+66.7% relative gain over dense search), an MRR of 1.0000, and 100% Historical Accuracy with a median query latency of 11.47 ms within a single local database file (memory.db).