Long-Term Memory
for AGI.

Production deployments fail because flat retrieval is version-blind. PaperGuru introduces a Lifecycle-Aware Memory that routes over citation graphs, respects deprecations, and scales infinitely.

Read Technical PaperView on GitHub
PaperGuru Platform Architecture — 66.05% +30.21% Lift
PaperBench

Clearing the human bar.

On PaperBench (OpenAI, 2025) — the toughest audit of an agent's ability to replicate ML papers from a PDF — prior systems fail to recover canonical conventions. By routing over the citation graph, PaperGuru achieves unprecedented performance.

66.05%
Mean score across all 23 papers
+30.21%
Absolute lift over the best published AiScientist baseline run on the shared subset.
19 / 20
Papers where PaperGuru beats the strongest published baseline.

Performance vs. Published Baselines

41% Human Expert (48h)
BasicAgent
21.0%
IterativeAgent
26.0%
AiScientist (Apr 2026)
33.7%
PaperGuru (ours)
66.0%
SurveyBench

AGI-scale writing.

On SurveyBench (150K+ tokens per output), PaperGuru generates surveys that score 94.66% on content under strict LLM judges. Because its evidence cards carry version-consistent provenance, it achieves a composite-richness lift where other agents score zero.

94.66%
Content Score (claude-opus-4.7 judge at 200K cap)
+14.06%
Absolute content lift over OpenAI DeepResearch.
+23.40%
Composite richness lift (100 verified figures, 200+ tables).

20 Generated Surveys

Click to open full PDF
3D Gaussian Splatting
3D Gaussian Splatting
3D Object Detection
3D Object Detection
Agentic RL
Agentic RL
LLM Alignment
LLM Alignment
Efficient Inference
Efficient Inference
LLM Evaluation
LLM Evaluation
LLM Explainability
LLM Explainability
Diffusion Models
Diffusion Models
Graph Neural Networks
Graph Neural Networks
LLM Hallucination
LLM Hallucination
LLMs for RecSys
LLMs for RecSys
LLMs for Time Series
LLMs for Time Series
Multi-Agent LLMs
Multi-Agent LLMs
Multimodal LLMs
Multimodal LLMs
RL for LLMs
RL for LLMs
RAG for LLMs
RAG for LLMs
LLM Safety
LLM Safety
Scientific LLMs
Scientific LLMs
Vision-Language-Action
Vision-Language-Action
Vision Transformers
Vision Transformers

Inside the Engine.

PaperGuru is powered by the Capital-Chunk Memory (CCM) kernel. It replaces flat similarity search with a temporal artifact graph, decoupling routing cost from archive size.

Formalism
A 4-axiom Lifecycle-Aware Memory framework mathematically proving that version-blind retrieval must accumulate stale evidence.
Kernel
Capital-Chunk Memory separates a bounded routing surface (compact heads) from unbounded payloads, achieving O(log N) routing complexity.
Real-world
End-to-end deployment study covering 30+ manuscripts, with 10 peer-reviewed acceptances (FSE, ICML, ICoGB, AEI).
Download Technical PaperGitHub Repository
PaperGuru Technical Paper Cover