I build agents
that remember.

Research and systems for agent memory, attention, and reliable computer use. First author of Proced-Mem, a benchmark for procedural memory retrieval in language agents (ICLR 2026 MemAgents Workshop). Currently building CASE.

GitHub X Scholar

Selected Work

CASE

A verified-procedure runtime that gives agents reliable hands inside real macOS apps. Typed, accessibility-targeted actions instead of pixel clicking: about 99% success where zero-shot agents manage 70%.

Runtime2026

Cortex

One knowledge graph built from every chat you have with AI. Turns conversations across ChatGPT, Claude, and Perplexity into decisions, people, and facts that agents read back over MCP.

Product2026

Proced-Mem

The first benchmark to isolate procedural memory retrieval from task execution in language agents. Seven retrieval methods across two domains, judged by an LLM protocol validated against human annotations.

Research2026

Attention Throughput Optimizer

Linear attention written in Triton: 50x faster than PyTorch SDPA at 64K sequence length on an A100, with peak memory cut from 19GB to 189MB.

Kernels2025

Agent-Debugger

Chrome DevTools for AI agents. Time-travel debugging with fork, patch, and replay on real LLM runs.

DevTools2026

Modality-Adaptive Attention

A 140M-parameter vision-language model built from scratch. Routes attention by token type, linear for vision and softmax where grounding needs sharpness, for 7% lower perplexity.

Research2025

after-effects-mcp

An MCP server with 70+ tools that lets agents drive Adobe After Effects: compositions, layers, keyframes, expressions, and motion templates.

MCP2026

Research

MemAgents Workshop, ICLR 2026

Proced-Mem: Benchmarking Procedural Memory Retrieval in Language Agents Across Domains

Ishant Kohar and Aswanth Krishnan

The first benchmark to isolate procedural memory retrieval from task execution, spanning text-based household tasks (ALFWorld) and real computer-use environments (OSWorld). Evaluates seven retrieval methods across text, visual, and lexical modalities with an LLM-as-judge protocol validated against human annotations.

Embedding-based retrieval degrades 30 to 42% on novel contexts, a generalization cliff: encoders treat procedures as unordered bags of tokens and discard the temporal structure that transfer depends on.

OpenReview Code

Path

2025 – Apr 2026Applied AI ResearcherQpiAI

2025AI EngineerQpiAI

2024Research Intern, Quantum ComputingIBM

2021 – 2025B.Tech ECENSUT, Delhi

Contact

GitHub X LinkedIn Scholar Blog

I build agentsthat remember.