Recallr AI Benchmarks

Welcome to the public repository for benchmarking Recallr AI against other memory providers (Supermemory, Mem0) on the LongMemEval (Oracle) benchmark.

Benchmark Results

Overall Accuracy (Pass@k)

Provider	Strategy	Accuracy
Recallr AI	Auto Recall	488/500 (97.5%)
Recallr AI	Agentic	466/500 (93.2%)
Recallr AI	Low Latency	439/500 (87.8%)
Recallr AI	Balanced	428/500 (85.6%)
Mem0	Non Graph	313/500 (62.6%)
Mem0	Graph	311/500 (62.2%)
Supermemory	Default	159/500 (31.8%)

Note: The Recallr AI website reports an overall accuracy of 97.5%, measured using Auto Recall, a routing layer added after these per-strategy benchmarks were run. Auto Recall automatically selects the best strategy (Agentic, Balanced, or Low Latency) for each input query, combining the strengths of all three.

Latency Statistics (Seconds)

Provider	Strategy	Min	P25	Median	P95	Max
Recallr AI	Low Latency	0.234	0.265	0.299	0.408	0.750
Recallr AI	Balanced	1.032	1.132	1.198	1.575	3.548
Recallr AI	Agentic	5.125	6.194	6.997	8.619	20.095
Mem0	Non Graph	0.489	0.504	0.786	1.787	6.171
Mem0	Graph	0.697	0.746	0.961	2.692	10.458
Supermemory	Default	0.392	0.851	1.301	3.293	4.242

---

Detailed Breakdown by Question Type

Recallr AI

Question Type	Agentic	Balanced	Low Latency
Knowledge Update	92.3%	94.9%	97.4%
Multi-session	89.5%	91.0%	91.0%
Single-session Assistant	100.0%	26.8%	26.8%
Single-session Preference	100.0%	93.3%	96.7%
Single-session User	100.0%	95.7%	98.6%
Temporal Reasoning	89.5%	92.5%	97.0%

Mem0

Question Type	Non Graph	Graph
Knowledge Update	76.9%	75.6%
Multi-session	65.4%	63.2%
Single-session Assistant	19.6%	19.6%
Single-session Preference	90.0%	90.0%
Single-session User	90.0%	90.0%
Temporal Reasoning	48.9%	50.4%

Supermemory

Question Type	Default
Knowledge Update	60.3%
Multi-session	35.3%
Single-session Assistant	3.6%
Single-session Preference	20.0%
Single-session User	30.0%
Temporal Reasoning	27.1%

---

Running the Benchmarks

Below are the commands used to run and evaluate each of the benchmark scripts on 500 records from longmemeval_oracle.json.

1. Recallr AI

Run the benchmark:

uv run python3 run_recallr_longmemeval.py \
    --data-path data/longmemeval/longmemeval_oracle.json \
    --start-index 0 --end-index 499 \
    --parallelism 20 --output-dir runs

Evaluate the results:

uv run python3 evaluate_runs.py \
    --provider recallr \
    --benchmark-version oracle \
    --requests-per-minute 200

2. Mem0

Run the benchmark:

uv run python3 run_mem0_longmemeval.py \
    --data-path data/longmemeval/longmemeval_oracle.json \
    --start-index 0 --end-index 499 \
    --parallelism 20 --output-dir runs

Evaluate the results:

uv run python3 evaluate_runs.py \
    --provider mem0 \
    --benchmark-version oracle \
    --requests-per-minute 200

3. Supermemory

Run the benchmark:

uv run python3 run_supermemory_longmemeval.py \
    --data-path data/longmemeval/longmemeval_oracle.json \
    --start-index 0 --end-index 499 \
    --parallelism 20 --output-dir runs

Evaluate the results:

uv run python3 evaluate_runs.py \
    --provider supermemory \
    --benchmark-version oracle \
    --requests-per-minute 200

---

Limitations

• These results use the LongMemEval Oracle dataset.
• The evaluation is produced and published by Recallr AI, not by an independent third party.
• Cloud provider behaviour may change after the recorded benchmark date.
• LLM-based evaluation can exhibit model-dependent judgment variance.

---

Reproducibility

We welcome independent reproduction, corrections, new provider integrations, and methodology improvements.

---

Contributing

Contributions are welcome! If you want to add new memory providers, datasets, or optimize existing strategies, feel free to open a pull request or submit an issue.