The case for confidence-weighted belief systems over vector similarity as the foundation for how AI agents remember.
Harshit Kumar
Backend engineer focused on AI memory infrastructure and distributed systems. Three years building Go services in fintech. Currently at Finplify, building Engram on the side.
Bengaluru, India · Open to remote (UK / EU / US)
Observation
Confidence Engine
decay · reinforce · contradict
Belief Store
PostgreSQL + pgvector
Agent Query
Engram
Most agent frameworks bolt on memory as an afterthought — a vector store that does nearest-neighbor search and calls it recall. Engram takes a different position: memory should model belief. Each fact carries a confidence score, decays over time unless reinforced, and can be contradicted or superseded by new observations. The goal is agents that reason about certainty, not just agents that retrieve context. Built in Go with PostgreSQL and pgvector.
The hypothesis: vector similarity is the wrong primitive for agent memory. Retrieval is not the same as belief.Read the deep-dive →
How RTMP ingest and WebRTC delivery work together in a low-latency streaming pipeline, and the architectural tradeoffs that come with sub-second latency constraints.
I design and own the Go-based ledger and reconciliation infrastructure — event-driven services, PostgreSQL, strict financial correctness requirements. Financial systems have zero tolerance for data errors; this role has pushed my engineering standards in ways that fast-moving product work rarely does. Currently the primary backend engineer on the reconciliation platform.
Built backend systems in Go and Node.js across multiple client product teams — APIs, data pipelines, early-stage product backends. First professional context where I owned systems end-to-end, from architecture decisions to production support. Shipped several products from zero to launch.
Engram
Cognitive memory system for AI agents with belief dynamics, temporal decay, and semantic retrieval. Covered in detail above.
DNS Server
A recursive DNS resolver written in Go with DNS-over-HTTPS (DoH) support and ad-blocking. Handles full query parsing, recursive traversal from root to authoritative name servers, and in-memory caching. Not a client wrapper — it implements the actual resolution algorithm. Includes a companion CLI for querying DNS records.
Live Streaming Service
Low-latency video streaming platform with WebRTC for real-time broadcast and RTMP for stream ingestion, with a Next.js frontend. The interesting work is in the signaling architecture and the latency tradeoffs — WebRTC’s sub-second delivery versus HLS’s reliability. Handles media routing, session management, and adaptive playback.
Glasskube
Kubernetes package manager. Added CLI features and fixed bugs in the installation wizard.
Distr
Developer distribution tooling by the Glasskube team. Infrastructure and deployment workflow contributions.
Hive
Infrastructure tooling contributions.
Stack Auth
Building the Go SDK alongside the core team for this open-source Clerk/Auth0 alternative.