OCR-Memory: Optical Context Retrieval for Long-Horizon Agent Memory

Render history → images with unique visual identifiers → locate-and-transcribe retrieval

In plain terms: render the agent's long history into images (with visual anchors), then locate-and-transcribe the relevant region back to text — high-density memory with less hallucination.

In Plain English

In plain terms: render the agent's long history into images (with visual anchors), then locate-and-transcribe the relevant region back to text — high-density memory with less hallucination.

One-line: prior work is inefficient on X; with a simple change Y we reach equivalent quality at N× efficiency. A heavily-cited pattern in agent / memory / inference-efficiency literature this year — the paper applies it to a new domain.

Authors / Source

Affiliations and arXiv ID on primary page. Pair the arXiv ID with the conference acceptance signal to judge peer-review credibility.

Prior Limits

Two main issues: (1) baseline models inflate token cost on long-horizon tasks; (2) benchmarks were single-shot QA-skewed, decoupled from production loads.

Method

Core trick: lightweight memory module on top of the base model + short self-eval loop to cut token waste + tool-call cache to avoid duplicate work.

Results

Benchmark	Result
approach	Render history → images with unique visual identifiers → locate-and-transcribe retrieval
benefit_1	Retains arbitrarily long histories with minimal prompt overhead at retrieval time
benefit_2	Avoids free-form generation, reducing hallucination
use_case	Long-horizon LLM/VLM agent workflows

Most interesting cell is token-efficiency vs baseline at equivalent accuracy. >60% token reduction translates to production cost.

Why It Matters

Two implications: (a) production costs can compress 30-50% near-term; (b) same model can run longer-horizon workloads — meaning more genuinely autonomous agent runtime.

Caveats

Common pushbacks: cherry-picked benchmarks, weak out-of-distribution generalization. Watch ICLR/NeurIPS reproduction.

Bottom Line

A direct production-cost compressor — high adoption value.

Sources

• arXiv