I build production LLM agent systems. Now focused on agent memory.

The question I'm working on: how agents consolidate, supersede, and forget what they know across sessions, and how to tell whether that memory is actually being used well over time, not just retrieved.

I write about it at Engram, working through the open problems in the field (consolidation, drift, evaluation) from first principles.

MemLayer - a memory-consolidation experiment for agents. A dual-buffer write path with supersession and archival, plus a benchmark that measures cross-session coherence instead of one-shot retrieval accuracy. (early, in progress)

Mostly closed-source, so the short version:

• Multi-agent orchestration over MCP and LangChain, with an observability layer for grounding, latency, and token metrics.
• A GraphRAG bond recommendation engine with hybrid retrieval: graph traversal plus dense vectors over LanceDB and Milvus.
• A MemGPT-inspired tiered-memory chatbot (profile / interaction / archival).
• LLM-as-judge evaluation pipelines built with DSPy.

Data engineering at TCS, and neuroimaging research at Rutgers (DTI, tractography). The interest in how memory consolidates traces back to that.

Fun Fact: I watch anime, sci-fi, and horror movies when I’m not deep in code.