Himanshu JhaAgentic AI · Memory · Backend
BuilderDevRel
6× hackathon winner

Himanshu
Jha.

AI agent developer.
Memory systems & backends.

I build agents that use tools, remember context and ask before they act. Here’s the software behind that work.

Explore my work I also explain what I build.
Native AI workspace

OmniDev

Play a real recording

Your context. An agent that can use it.

Watch the Mac app inspect files, propose an edit and ask before running a command.

Native app recording · no narration · waiting time acceleratedInside the OmniDev build
OmniDev / Agent in action
OmniDev native agent workspace with a conversation and visible tool activity
Native app recording · no narration · waiting time acceleratedOpen original image
Memory & contextMCP & agent systemsBackend engineeringSelected work
Selected builds / See them in use

Less pitch. More product.

From the memory an agent trusts to the tools it uses and the work it saves. Explore the actual interfaces here.

Agent memory

ContextFirewall

Play a real recording

See what gets remembered. And what gets blocked.

A recorded console walkthrough: memory graph, audited recall, reusable rules and forgetting.

Recorded sample project · on-screen explanations · no audioInside the ContextFirewall build
ContextFirewall / Watch the lifecycle
ContextFirewall console walkthrough of the memory lifecycle
Recorded sample project · on-screen explanations · no audioOpen original image
WebMCP tooling

Graft

Explore the product

A website becomes an inspectable tool.

Compare the source page with its generated contracts, then inspect the evidence behind a tool.

Browser capture · owned demonstration catalogInside the Graft build
Graft / Compile a website
Graft workbench showing the Signal Cabinet catalog beside derived WebMCP tools and a contract inspector
Browser capture · owned demonstration catalogOpen original image
Voice workflows

VOX

Explore the product

A spoken request becomes reviewable work.

Explore the saved run, sandbox workspace and mobile approval queue. The tickets are fictional.

Repository capture · historical run · fictional ticketsInside the VOX build
VOX / A completed run
VOX completed run with original instruction, 24 tickets triaged, seven saved labels and two prepared drafts
Repository capture · historical run · fictional ticketsOpen original image
Explore all four case studies

The part after
the prompt.

I’m Himanshu Jha, also known as Himanshu Kumar and himanshu748.

Carrying context across sessions. Using a tool safely. Recovering from a failed action. I build across Python and TypeScript, APIs and data stores, MCP tools and user-facing products.

Memory that can changeProvenance, revisions, forgetting, and retrieval.

Actions with boundariesScoped access, approval gates, and auditable decisions.

Backends that keep their promisesPersistent state, retry-safe writes, and observable workflows.

Memory, with a history.

The deployment command is stored with the instruction it came from.

Original instructionCurrent
flyctl deploySession 01 · instruction 01
Revised instructionPending
make releaseSession 01 · instruction 02
Next agent sessionWaiting
—Session 02 · ready to recall

Instruction 01 is current. Its source stays attached.

Inspect the record
Current value
flyctl deploy
Source
Session 01 · instruction 01
History
First revision. No earlier value.
Retrieval
No retrieval requested.

Interactive illustration · no model calls

Selected writing / Five deep dives

Build it.
Break it. Explain it.

Agent memory, evaluation and observability—explained through actual failures, source code, and the decisions that followed.

Agent memory / Graphs

EmbodiWeave: my agent read its own failure and repeated it, so I made the graph enforce it (FalkorDB + MCP)

A recorded failure changed the design: graph-enforced decisions, fixed promotion gates and a robot simulation that could refuse a regression.

Agent evaluation / Trust

ToolTrap: My support agent handed a customer a fake phone number. The fix worked, until the fake moved.

Testing fake-detail leakage, legitimate-detail retention and the limits of prompt-only defenses. Includes scorer corrections and held-out cases.

Observability / Engineering

We instrumented an AI agent swarm with SigNoz, and its own telemetry told us we were wrong about almost everything

Tracing a multi-agent system exposed mistaken assumptions about routing, budgets and review quality. The article follows the telemetry to the fix.

Memory / Architecture

I built a trust firewall for my AI agent's memory — on Cognee's four verbs

From Cognee’s memory lifecycle to a reviewed context pack: provenance, conflicting facts, embedding failures and the checks around recall.

Debugging / Open source

The smolagents bug that made my agent retry the same valid code three times

A minimal reproduction, an interpreter bug and a fix explained step by step. Recognized in DEV’s Summer Bug Smash runner-up announcement.

All writing on DEV

Have a hard
problem?

Agentic AI, memory infrastructure, or a backend problem you’re working through. Tell me what you have in mind.