Long-horizon memory: survey of seven architectures, ranked by recall and cost
Compares episodic, semantic, hybrid, and graph-based memory across realistic 30-day agent simulations. Hybrid stores win on recall; graph stores win on cost stability.
A 30-day simulated deployment compares seven memory architectures across recall, latency, and amortized cost. Hybrid stores (episodic + semantic + summary) lead recall by 12 points but cost 2.4× more than graph-based stores at month three.
What changed. First like-for-like comparison of memory architectures over a long enough horizon to surface compaction and decay behavior.
Why it matters. Memory is where agent quality silently degrades over weeks. Choosing the wrong store at month one can quietly compound until users churn at month three.
Builder takeaway. If you have a hot retrieval path with high QPS, a graph-backed store is hard to beat. If you have rare but high-stakes recall (legal, medical, executive assistant), pay for the hybrid.