Reliability you can reproduce

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.

How to read this: these are asymmetric proofs — they run CLS++ only. We do not execute competitors; we link their open issues, which document the failure classes. No accuracy leaderboard, no “#1” claims — just whether the failure can happen in CLS++ (it can't) and a command you can run.

Silent memory loss when the embedder fails

mem0 #5245

The 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.

0% loss — 100% of facts retained under a fully-broken embedder.

Delete-on-conflict — a contradiction wipes the old memory

mem0 #4536

The reconsolidation gate archives the superseded fact with full lineage and version, and surfaces the new one. Nothing is deleted; the change history is provable.

Old fact archived with lineage, not deleted — history preserved.

Silent write loss on an embedding-dimension change

mem0 #4985

An embedding-dimension mismatch is non-fatal — the item is persisted regardless and reads back.

Writes survive a dimension switch — read back intact.

Empty extraction returns success with nothing stored

mem0 #3009

The raw item is always persisted — there is no LLM-extraction step that can silently drop a valid input.

Every write is stored — never a silent success-with-nothing.

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