EXIA GHOST Results: 89.94% (Cat 1-4), first Cat 5 evaluation (71.52%)

[francisdu53]

Hello LoCoMo Team,

We are submitting results for EXIA GHOST V5, a bio-inspired AI memory system with 5 specialized stores (contextual, semantic, episodic, procedural, prospective).

Results (Strict Judge)

Category	Accuracy
Cat 1 — Single-hop	93.69%
Cat 2 — Multi-hop	86.30%
Cat 3 — Open-ended	91.43%
Cat 4 — Temporal	88.36%
Cat 5 — Adversarial	71.52%
Average (1-4)	89.94%
Average (1-5)	85.80%

Notable: this is the first public evaluation including Category 5 (adversarial/unanswerable questions).

Verification

• Repository: https://github.com/francisdu53/exia-ghost-benchmarks
• Full results & methodology: locomo/RESULTS.md
• Judge: GPT-4o-mini (strict, CORRECT/WRONG), same standard used by Mem0 and MemMachine
• All 10 conversations, 20 speakers, 1,339 QA pairs

Author: Francis BABIN (Solo Developer)

We appreciate your work on this crucial benchmark. Thank you.

[francisdu53]

Update — Judge Robustness Validation

We re-ran the full benchmark using GPT-4o-mini as both QA generator and judge (replacing Claude Haiku 4.5 for QA generation).

Category	Haiku QA	GPT-4o-mini QA	Delta
Cat 1 — Multi-hop	86.17%	86.52%	+0.35
Cat 2 — Temporal	85.36%	88.16%	+2.80
Cat 3 — World knowledge	93.75%	94.79%	+1.04
Cat 4 — Single-hop	92.51%	91.56%	-0.95
Cat 5 — Adversarial	71.52%	72.42%	+0.90
Overall (1-5)	85.80%	86.15%	+0.35

Scores are consistent across LLM providers (max delta: 2.80%). The memory architecture drives the score, not the LLM choice.

Full results added to the repository: locomo/RESULTS.md