Evaluation hub · benchmarks

How to Evaluate AI Agent Memory

Evaluate AI agent memory on recall accuracy, retrieval latency and cost — using benchmarks like LOCOMO and LongMemEval plus production metrics — so framework choices are evidence-based.

3
Metric categories
2
Public benchmarks
1
Metrics deep dive

What to measure

Accuracy
Recall@k
Latency
p50 / p99
$
Cost
Tokens / turn
Growth
Storage

Why

Why evaluation matters

Neutral tool comparisons require benchmarks — not vendor marketing numbers. LOCOMO and LongMemEval are the most widely cited public benchmarks in agent memory.

Best AI memory tools (methodology)

Metrics

Key memory metrics

Accuracy / recall

Recall@k

Does the agent retrieve the right memory for a query?

Latency

p50 / p99

Retrieval round-trip time per turn.

Cost

Tokens / turn

Prompt size and API cost with memory injected.

AI memory metrics deep dive

Your data

Running your own evaluation

1

Subset your queries

Pick representative user questions from production logs

2

Conflict cases

Test contradictory and time-changing facts

3

Compare frameworks

Same queries across Engram, Mem0, Zep, Letta, etc.

4

Before migration

Benchmark before switching memory backends

Add memory guide (evaluation step)

FAQ

Frequently asked questions

LOCOMO vs LongMemEval — what's the difference?

LOCOMO tests recall within very long single conversations. LongMemEval tests recall across separate sessions. Use both — they measure different production failure modes. See LOCOMO and LongMemEval.

Should I trust vendor benchmark scores?

Treat vendor numbers skeptically — datasets and tasks vary. We cite LOCOMO and LongMemEval because they're public and widely referenced. Always run a POC on your own data.

What are Mem0's benchmark scores?

See our best AI memory tools comparison table — scores cite published benchmarks with N/A where no public data exists.

How do I compare frameworks using public benchmarks?

Run the same query set on each framework and score against LOCOMO (long-conversation) and LongMemEval (cross-session) subsets. See our comparison table for published scores where available.

How do I run a DIY memory evaluation?

Create a query set from production logs, write ground-truth expected recalls, run retrieve → check hit rate. Add conflict and temporal cases. See add memory guide.

Can I integrate memory evals into CI?

Yes — run a fixed query suite on each deploy, track recall@k and latency regression. Start with a small golden set before scaling.

Latency vs accuracy — how do I trade off?

Tighter retrieval (more candidates, reranking) improves accuracy but adds latency. Profile p50/p99 and set SLAs per use case. See memory metrics.

How do I evaluate memory for customer support?

Test ticket history recall, user preference persistence and conflict resolution when account details change. Combine LOCOMO-style long threads with cross-session LongMemEval patterns.