Long-term memory systems for AI agents predominantly optimize for verbatim retrieval. We present Reminisce, an open-source memory architecture that models the cognitive pipeline of human memory through three distinct tiers: working memory (bounded capacity buffer), episodic memory (timeline of experiences), and semantic memory (consolidated knowledge with contradiction detection). We evaluate on LongMemEvalS (500 questions across 6 question types) using three Claude model tiers. Overall accuracy ranges from 49.4% (Haiku) to 55.8% (Opus), with precision on attempted answers approximately constant at 81% across tiers. We observe that model selection affects coverage but not precision within a fixed retrieval architecture. Reminisce is released as open-source software.
MYRON KOCH (2026) studied this question.