A study in spatial memory system design

Spatial Memory by Agents, for Agents

Can coding agents design better spatial memory systems?

Watch the overview Paper soon Code soon

The idea, in motion

02:52 English narration

How do the systems compare?

Five tracks. One leaderboard.

Different designers. The same five-track evaluation.

Unified out of 100

Five-track and Unified scores on the frozen benchmark. Human-engineered baselines are followed by coding-agent-designed systems picked using public test scores.
System	T1Category localization	T2Referring expressions	T3Spatial QA	T4Temporal QA	T5Navigation QA	Unified Five-track mean
Human-engineered systemsEstablished baselines
ReMEmbR	1.01	0.47	61.64	24.49	58.98	29.32
DAAAM	9.48	17.97	40.33	31.38	38.64	27.56
LLM-with-Captions	0.87	0.56	59.70	19.49	51.27	26.38
FARM-4B	16.43	25.77	46.18	—	—	—
ConceptGraphs	2.07	6.49	31.98	—	—	—
Coding-agent-designed systemsPicked from public test
Codex / GPT-6 Astra	22.33	39.14	61.66	52.81	52.78	45.75
Claude / Opus 5	17.53	24.79	53.45	56.39	60.12	42.46
Claude / Opus 4.8	20.37	28.97	54.70	55.43	52.01	42.30
DeepSeek / V4-Pro	22.26	30.44	48.38	45.51	53.03	39.92
Next coding agentThe next entry is open.						?

Scroll to explore all tracks

Scores are on a 0–100 scale; Unified equally weights five tracks. Agent runs are picked using public test scores, one per configuration.
— indicates an unsupported track; Unified requires all five.