Anup Roy
Research Subjects
Anup Roy has a total of 5 co-authors. Here are the top co-authors by number of shared publications:
- Aidan Philip Millar
Mubadala Investment Company, Abu Dhabi, United Arab Emirates
2 shared publications - Larry Murray
Inception42, Abu Dhabi, United Arab Emirates
2 shared publications - Rishabh Upadhyay
University of Milano-Bicocca
2 shared publications - Robin Mills
Inception42, Abu Dhabi, United Arab Emirates
2 shared publications - Animesh Rameshbhai Panara
Inception AI, Abu Dhabi, United Arab Emirates
1 shared publication
Latest Publications
- Open Access
- 6 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 7726-7731
PyRAMIND: Auditable Deterministic Pyramid Memory for Long-Running LLM Agents
Long-running LLM agents are increasingly considered for regulated, high-consequence domains such as finance, compliance, and audit, where memory must be accurate, traceable, replayable, and inspectable after a decision. Existing agent-memory systems often invoke an LLM during maintenance operations such as fact extraction, consolidation, updating, summarization, and forgetting. Although effective, this design makes it harder to reconstruct why a memory existed, changed, or was removed at a particular point in time. We present PyRAMIND (Pyramidal Retention with Arithmetic Memory for Inference-time Determinism), an auditable memory architecture in which write consolidation, reinforcement, forgetting, tier promotion, and retrieval scoring are closed-form arithmetic over cached embeddings. Only optional entity tagging at ingest and answer generation at query time call an external LLM. Given fixed embeddings, cached boundary-model outputs, parameters, and software, each maintenance transition is replayable from ordered inputs. PyRAMIND stores memories in a four-tier Pyramid (raw →facts →themes→worldview) and ranks them with a max-ensemble retention score modulated by local semantic density. Under a matched harness, PyRAMIND achieves competitive accuracy on long-horizon memory tasks (0.93 on LongMemEval-S) despite eliminating LLM calls from the memory-maintenance path, while reducing memory-lifecycle token expenditure to 1.4k LLM tokens per turn versus 3.1k for reproduced Mem0 OSS and 4.2k for Zep v3. Performance is weaker on temporal and multi-hop reasoning tasks (0.62 on LoCoMo), highlighting the continued value of richer relational memory structures for compositional reasoning. These results suggest that strong long-term memory performance is possible without LLM-mediated maintenance, and motivate future replayable memory systems that add temporal and relational structure while preserving auditable deterministic maintenance.
- 6 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 7726-7731
Open AccessPyRAMIND: Auditable Deterministic Pyramid Memory for Long-Running LLM Agents
Long-running LLM agents are increasingly considered for regulated, high-consequence domains such as finance, compliance, and audit, where memory must be accurate, traceable, replayable, and inspectable after a decision. Existing agent-memory systems often invoke an LLM during maintenance operations such as fact extraction, consolidation, updating, summarization, and forgetting. Although effective, this design makes it harder to reconstruct why a memory existed, changed, or was removed at a particular point in time. We present PyRAMIND (Pyramidal Retention with Arithmetic Memory for Inference-time Determinism), an auditable memory architecture in which write consolidation, reinforcement, forgetting, tier promotion, and retrieval scoring are closed-form arithmetic over cached embeddings. Only optional entity tagging at ingest and answer generation at query time call an external LLM. Given fixed embeddings, cached boundary-model outputs, parameters, and software, each maintenance transition is replayable from ordered inputs. PyRAMIND stores memories in a four-tier Pyramid (raw →facts →themes→worldview) and ranks them with a max-ensemble retention score modulated by local semantic density. Under a matched harness, PyRAMIND achieves competitive accuracy on long-horizon memory tasks (0.93 on LongMemEval-S) despite eliminating LLM calls from the memory-maintenance path, while reducing memory-lifecycle token expenditure to 1.4k LLM tokens per turn versus 3.1k for reproduced Mem0 OSS and 4.2k for Zep v3. Performance is weaker on temporal and multi-hop reasoning tasks (0.62 on LoCoMo), highlighting the continued value of richer relational memory structures for compositional reasoning. These results suggest that strong long-term memory performance is possible without LLM-mediated maintenance, and motivate future replayable memory systems that add temporal and relational structure while preserving auditable deterministic maintenance.
- Open Access
- 19 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 4087-4092
BANANA: Bounded Adaptive Navigation Architecture for Nested Archives
- 19 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 4087-4092
Open AccessBANANA: Bounded Adaptive Navigation Architecture for Nested Archives