LLM agents do not persist: they forget across sessions, and within a growing memory they surface stale facts as if current. We present Continuum, a tiered memory system whose long-term store is bi-temporal — every fact carries a valid-time (when it held in the world) and a transaction-time (when the system learned it) — with soft supersession that retires a contradicted fact in place, keeping retrieval current while history stays intact and auditable. On scripted knowledge-update and point-in-time ("as-of") benchmarks, Continuum answers 100% correctly, versus 38% for an append-only store and 20%/75% for single-axis baselines. Run on the same scenarios, the real Mem0 SDK — which resolves contradictions by deleting the superseded fact and has no valid-time axis — is competitive at returning the latest value (72%) but scores 27% on point-in-time queries and 30% on the bi-temporal set overall. The gap is structural: deleting a fact destroys the history an as-of query needs. On LongMemEval-S Continuum reaches ~74% judged accuracy; a controlled decomposition finds the ceiling is answerer reasoning, not retrieval. Continuum is open-source, hybrid (bge-m3 + BM25 fused with RRF over an HNSW-indexed pgvector store), and MCP-exposed; we release the code and benchmark sets.
Mayank Sahu (2026) studied this question.