daven3/memorycraft-bench

MemoryCraft — Unified Agent-Memory Benchmark Collection

Five memory benchmarks reformatted into one common schema for evaluating how well an agent uses long-term memory. Two configs:

• full/ — every instance of each source, unified.
• selected/ — the evaluation subset used in our runs (QA balanced across benchmarks; Membench = its largest/long-context instances).

benchmark	full instances	full QA	selected instances	selected QA
locomo	10	1986	10	994
longmemeval	500	500	500	500
memoryagentbench	146	3671	18	1509
ama_bench	208	2496	125	1500
membench	23166	23166	1158	1158
Total	24030	31819	1811	5661

Unified schema (one JSON object per line)

{
  "uid": "<source>__<id>", "source": "...", "memory_type": "conversation|document|trajectory",
  "sessions": [{ "session_id","timestamp","turns":[{"speaker","role","content","turn_id","metadata"}] }],
  "qa": [{ "qa_id","question","answer","question_type","raw_category","evidence","abstention","metadata" }],
  "metadata": { }
}

Sources & licenses

Reformatted redistribution — please cite and follow each original license:

• LoCoMo — snap-research/locomo
• LongMemEval — xiaowu0162/longmemeval-cleaned
• MemoryAgentBench — ai-hyz/MemoryAgentBench
• AMA-bench — AMA-bench/AMA-bench
• Membench — import-myself/Membench

Repo: daven3/memorycraft-bench. Reformatting + sampling scripts live in data_prep/ and eval/.