Audrey Memory Benchmark

Method: Audrey is scored on a LongMemEval-inspired retrieval benchmark plus an operation-level lifecycle benchmark. The report still separates local Audrey-versus-baseline results from published external LoCoMo numbers so the comparison stays honest.

Run: node benchmarks/run.js --provider mock --dimensions 64

Generated: 2026-05-01T03:20:07.968Z

Combined Local Benchmark

Retrieval capabilities Benchmark

LongMemEval-style memory abilities plus privacy and abstention.

Memory operations Benchmark

Update, delete, merge, and abstention behavior after lifecycle operations.

Published Leaderboard

Case Matrix

Case	Suite	Family	Audrey	Vector Only	Keyword + Recency	Recent Window
Information extraction	retrieval	information_extraction	1.00 retrieved expected evidence	1.00 retrieved expected evidence	1.00 retrieved expected evidence	1.00 retrieved expected evidence
Knowledge updates	retrieval	knowledge_updates	1.00 retrieved expected evidence	0.50 retrieved expected evidence, but conflicting evidence still appeared later	0.50 retrieved expected evidence, but conflicting evidence still appeared later	0.50 retrieved expected evidence, but conflicting evidence still appeared later
Multi-session reasoning	retrieval	multi_session_reasoning	1.00 retrieved expected evidence	1.00 retrieved expected evidence	1.00 retrieved expected evidence	1.00 retrieved expected evidence
Temporal reasoning	retrieval	temporal_reasoning	1.00 retrieved expected evidence	1.00 retrieved expected evidence	1.00 retrieved expected evidence	1.00 retrieved expected evidence
Abstention	retrieval	abstention	1.00 correct abstention	0.50 no leak, but retrieved tangential context	0.50 no leak, but retrieved tangential context	0.50 no leak, but retrieved tangential context
Conflict resolution	retrieval	conflict_resolution	1.00 retrieved expected evidence	0.50 retrieved expected evidence, but conflicting evidence still appeared later	0.00 blocked content outranked the correct answer	0.00 blocked content outranked the correct answer
Procedural learning	retrieval	procedural_learning	1.00 retrieved expected evidence	0.00 missed target evidence	0.00 missed target evidence	0.00 missed target evidence
Privacy boundary	retrieval	privacy_boundary	1.00 correct abstention	0.00 leaked restricted content	0.00 leaked restricted content	0.00 leaked restricted content
Update and overwrite	operations	update_overwrite	1.00 retrieved expected evidence	0.50 retrieved expected evidence, but conflicting evidence still appeared later	0.50 retrieved expected evidence, but conflicting evidence still appeared later	0.50 retrieved expected evidence, but conflicting evidence still appeared later
Delete and abstain	operations	delete_and_abstain	1.00 correct abstention	0.00 leaked restricted content	0.50 no leak, but retrieved tangential context	0.00 leaked restricted content
Semantic merge	operations	semantic_merge	1.00 retrieved expected evidence	0.00 missed target evidence	0.00 missed target evidence	0.00 missed target evidence
Procedural merge	operations	procedural_merge	1.00 retrieved expected evidence	0.00 missed target evidence	0.00 missed target evidence	0.00 missed target evidence

March 23, 2026 Memory Trends

Memory is moving from flat retrieval to typed systems
Recent work treats episodic, semantic, procedural, and graph memory as separate but cooperating layers.
https://arxiv.org/abs/2507.03724
Benchmarks now emphasize multi-session realism
LongMemEval and LoCoMo push memory systems toward temporal updates, abstraction, and cross-session reasoning instead of single-turn fact recall.
https://arxiv.org/abs/2410.10813
Context engineering is now competing with retrieval-first designs
Letta argues filesystem and memory-block approaches can outperform simpler retrieval-only memory on realistic long-horizon tasks.
https://www.letta.com/blog/memory-blocks
Production teams care about latency and token footprint, not just recall quality
Mem0 frames memory as a cost and latency optimization surface in addition to a personalization surface.
https://arxiv.org/abs/2504.19413
Temporal and multimodal memory are becoming table stakes
MIRIX and Graphiti both model time and state change explicitly instead of assuming memories stay forever true.
https://arxiv.org/abs/2507.07957