Memory that doesn't silently lose or wrongly overwrite a fact
Below are runnable reproductions of failure classes documented in the leading memory layer's own open GitHub issues — and how CLS++ avoids each one. Don't take our word for it; run them yourself.
Silent memory loss when the embedder fails
mem0 #5245The fact persists from the engine's own tokenization; dense embeddings are an optional re-rank attached after the write. A dead embedder degrades ranking — it loses nothing.
Delete-on-conflict — a contradiction wipes the old memory
mem0 #4536The reconsolidation gate archives the superseded fact with full lineage and version, and surfaces the new one. Nothing is deleted; the change history is provable.
Silent write loss on an embedding-dimension change
mem0 #4985An embedding-dimension mismatch is non-fatal — the item is persisted regardless and reads back.
Empty extraction returns success with nothing stored
mem0 #3009The raw item is always persisted — there is no LLM-extraction step that can silently drop a valid input.
Durable by design
Writes are acknowledged only after an fsync to a write-ahead log, with replication off the hot path. The recovery path is tested against process death, torn-tail records, and both replicas down — so an accepted write is a kept write.
Run the proof
The benchmarks are CLS++-only and deterministic. Install the package and run:
pip install clsplusplus
cls bench # runs the reliability reproductions
cls bench --json # machine-readable summary