Product parity
TemporalStore and a leading memory library, feature for feature.
Hierarchical memory libraries made event-and-entity context memory intuitive, and it is a genuinely good idea. TemporalStore — the open-source engine behind MatrixArk — keeps that shape and turns it into a production serving layer: freshness, replay, filter-first retrieval, and bounded time windows. On a shared open-source harness with an LLM judge, that difference showed up as 88% vs 78% overall.
This is not a takedown. That approach validated something MatrixArk believes deeply: that context memory should be structured — events, entities, and a readable hierarchy — not a flat pile of embeddings. The question this page answers is narrower and more useful: where do the two overlap as products, and where does a serving-grade layer pull ahead?
The shared foundation
Both systems ingest raw conversation and tool activity, extract structured memory, and assemble a context pack for the model. Both organize memory hierarchically. Both support scoped recall. If your workload is a short session where the relevant fact is recent, they behave very similarly — which is exactly why they tie at 83–84% on LOCOMO.
Where a serving layer diverges
A leading memory library is exactly that — a library. TemporalStore is a memory library plus a serving engine.
| Capability | Leading memory library | TemporalStore |
|---|---|---|
| Hierarchical event/entity memory | Yes — the original strength | Yes — compiled into scope-hashed context nodes |
| Time-aware freshness | Limited; recency heuristics | First-class: validity windows and stale-memory blocking |
| Filter-first retrieval | Vector-led recall | Bounded time-window + filter traversal, vector optional |
| Replay / audit | Not a serving guarantee | Replayable context packs by id for evals and audit |
| Long-horizon recall | 66% on LongMemEval | 98% on LongMemEval |
| Retrieval hit@k (LongMemEval) | 0.88 | 1.00 |
| Token cost at equal quality | Baseline | Large savings |
| Deployment | Library | Open-source engine + enterprise platform & storage backends |
How the products fit together
The hierarchy is the shared idea; the serving path is what makes it production-ready.
Same hierarchy, but every read is bounded, time-aware, and replayable — which is what turns a good memory idea into infrastructure a vertical AI product can depend on.
The honest summary
If you want a memory library and your sessions are short, a leading memory library is a reasonable choice and the two systems land in the same place. If you are shipping a product where the model must recall the right fact from far back, ignore what has since changed, and let you replay exactly what it saw, the serving guarantees are the difference — and the benchmark reflects it: 88% vs 78% overall, 98% vs 66% where memory is hardest.