Research & Interests
What I keep coming back to, at work and off hours — and the papers that shaped how I think about it. Each paper gets a TL;DR and a note on how it showed up in something I built.
Topics I follow
- Private, user-owned memory for AI agents. If an agent remembers your life, the platform running it shouldn't be able to read that memory. VaultMem, daylog.
- Brain-inspired retrieval. Associative memory, attractor dynamics and consolidation as alternatives to plain vector search. Attractor.
- Local-first AI. Running useful models on hardware you own — transcription, embeddings and chat without sending data to third parties. Whisper on a $0 server, my setup.
- Distributed systems & API design at scale. My day job: aggregating data across shards and products into APIs that stay fast for the biggest customers. Experience.
- AI-assisted software engineering. Where AI genuinely speeds up engineering work — and how to keep its output honest. Doing the right thing, fast, reviewing AI-generated tests.
- Ruby & Rails architecture. Thirteen years in the community, from RubyConf India 2013 to SF Ruby 2025. Events.
My own work
- VaultMem: zero-knowledge encrypted memory for AI agents — research preprint, 2026. Zenodo (DOI) · PyPI · Code
Papershelf
Prefer video? See talks worth watching.
Agent memory & privacy
-
Memory as Asset: From Agent-centric to Human-centric Memory Management
TL;DR Argues that today's agent memory is built for the agent, not the person, and proposes treating personal memories as assets people own, share under permissions and evolve over time — with a three-layer design of personal storage, an evolution layer and a memory exchange network. Full write-up →
Why it matters to me The paper that shaped VaultMem's whole design. VaultMem implements its first layer: memory the platform operator cryptographically cannot read. Read more →
-
Evaluating Very Long-Term Conversational Memory of LLM Agents (LoCoMo)
TL;DR Introduces LoCoMo, a benchmark of very long multi-session conversations with question-answering across single-hop, multi-hop, temporal, adversarial and open-domain categories — and shows long-context LLMs and RAG still struggle with long-range memory. Full write-up →
Why it matters to me I used its 1,982 QA pairs to evaluate VaultMem's retrieval instead of my own small synthetic benchmark. Read more →
Brain-inspired memory
-
Neural networks and physical systems with emergent collective computational abilities
TL;DR A network of simple binary neurons with symmetric connections has an energy function that only goes down as it updates, so stored patterns become stable "attractors" — feed it a partial or noisy cue and it settles on the full memory. Content-addressable memory from physics. Full write-up →
Why it matters to me The starting point for Attractor, my memory engine that retrieves by partial cues and context instead of keys or nearest-neighbour distance. Read more →
-
Hopfield Networks is All You Need
TL;DR A continuous "modern" Hopfield network can store exponentially many patterns and retrieve them in one update — and that update rule is exactly transformer attention. Full write-up →
Why it matters to me The bridge between 1982 associative memory and today's LLMs, and the reason I think attractor-style retrieval is worth revisiting for agents. Read more →
-
Why there are complementary learning systems in the hippocampus and neocortex
TL;DR The brain pairs a fast learner (hippocampus) that captures specific episodes with a slow learner (neocortex) that gradually extracts structure; replaying episodes into the slow system avoids catastrophic interference. Full write-up →
Why it matters to me Gave Attractor its model of consolidation — why some memories strengthen with use while others decay. Read more →
Retrieval & embeddings
-
Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks
TL;DR Fine-tunes BERT in a siamese setup so each sentence gets its own embedding comparable by cosine similarity — turning semantic search over many sentences from hours of pairwise BERT inference into seconds. Full write-up →
Why it matters to me The sentence-transformers models behind semantic search in Attractor, VaultMem and daylog all descend from this paper. Read more →
Newsletters I read
- Every — essays on AI and how people actually work with it.
- SF Ruby and the Ruby Central README — what's happening in the Ruby community.
- Software Engineering Daily — infrastructure and engineering deep dives.
- We Love Open Source — a weekly roundup of open-source projects and people.
- Indie Hackers — people shipping small products on their own.
- ACM TechNews — a quick scan of computing research news.