Right where you left off.Sharper than yesterday.
Infinite conversation and coding in a single session — a companion agent that remembers everything from yesterday.
JLC no-prefix route active
MEM project state persisted across session
ZERO handoff · switching · onboarding ready
OMM mistake converted into project warning
RUN turn 9,842 · compact 0 · clear 0 · no ceiling
Every LLM forgets between turns — like Leonard in Memento.
Other agents cope by re-reading the whole conversation every turn. Push it to the limit and it loses everything it had remembered, all at once.
JARVIS CODE just writes the memory down outside and reads it back — like Leonard's notes. It carries the notes, not the whole past, so it never forgets and never slows down.
JLC is the harness that keeps those notes sharp — so Leonard never walks off in the wrong direction.
Notes by default. The diary when Leonard reaches back.
The JHB is Leonard's notes — it rides every turn. When your message reaches into the past, JLC also lets him flip back through the diary: it reconstructs the relevant raw pages on the fly — BM25 to narrow, bge‑m3 cosine to rerank, the top few handed over — never the whole history dumped back in. The notes carry the understanding; the diary backs them with the exact words, only when you reach for them. How it's built →
Every agent is actually
stateless.
An LLM remembers nothing between turns. Existing agents hide this by replaying the entire conversation every turn — fighting their own nature, until the context fills and collapses.
JARVIS CODE is built for statelessness. Working with the grain, not against it — it never slows as it grows, and never forgets.
You think you're raising an AI.
But every session,
you meet a stranger.
Every time you re-explain yesterday, you lose your most precious resource: time.
Every time it forgets, you're the one filling the gap.
The AI built to get smarter only gets dumber as the chat grows.
JARVIS CODE starts today with yesterday intact.
Your time goes to creating — not re-explaining.
Drags a finite session
/compacted — context compressed and lost/clear throws memory awayCarries infinite memory
The JARVIS LLM Codec
bypasses the Transformer's curse.
Attention is O(n²) — the longer the chat, the more cost balloons, until /compact·/clear throw memory away. JLC carries the memory outside the window, so the curve bends back to linear.
It forgets — so you never have to.
Linear cost, however long
No giant conversation prefix dragged along.
Cost per turn stays linear — O(n), not O(n²) — so long runs finish without a blowup.
Memory, not tied to a model
It lives in the codec, not the model.
Restart, swap models, or drop to a smaller one — your project context stays put.
No context ceiling
No compact, no clear. Spread across state and memory, there's no window left to fill — infinite conversation and coding in one session.
Light on chat. Deep on code.
Plain reasoning for everyday chat.
For coding it pushes reasoning to the limit to pull out the model's very best — and with memory carried outside, it does that without ever burning down the context.
The more you use it,
the smarter it gets about your codebase.
Self-improving memory
Day 3 of refactoring 200 files. You close the laptop Friday night, open it Monday morning. JARVIS CODE resumes right where it stopped — remembering even why you abandoned approach X.
OMM records its own mistakes.
Without you pointing them out, the same mistake never repeats — and the more you work on a project, the smarter it gets on its own.
Compounding you can't fake.
Curious how JARVIS CODE actually works? Watch the 48-minute video demo →
Your code is yours.
No servers. No telemetry.
jarvis-code has no backend. Nothing phones home — no analytics, no usage tracking, no update pings. It even disables the embedding model's own telemetry.
Bring your own key. Zero markup.
A solo dev can't run proxy servers — so it doesn't. You connect your own provider (Anthropic, OpenAI, Ollama…) and pay the vendor directly. jarvis-code just provides the codec and the memory harness.
Local, readable memory.
Your memory lives on your disk as plain-text JSONL — not a locked binary DB. Embeddings run locally (bge-m3). Your code only leaves for the LLM provider you choose.
First install downloads the ~2.3 GB embedding model once from HuggingFace; web search, if you enable it, uses Brave. That's the full list.
The coding companion that never forgets.
Get started now.
You may already use another agent. It's perfect at first, too.
The moment the chat grows and the context fills, it leaves you behind.
JARVIS CODE begins exactly there.
🤖 Already using an AI assistant? Let Claude Code, Codex, or Cursor do the whole install — it tailors the setup to the AI you already pay for. Paste it this: