Developer Preview

Stop paying AI
to re-read your codebase
every session.

Every new Copilot session starts blind. Your codebase gets re-sent, re-read, and re-familiarized before a single useful answer comes back — every morning, every developer, fresh bill. Semantic Cache Engine caches that context locally. Warm sessions answer in ~41ms with zero tokens used. When a teammate imports your bundle, they skip straight to warm.

75–78% session-start token reduction
51× faster on cache hits
0 outbound connections

We measured the waste. Then we stopped it.

Every new Copilot session starts blind. Same codebase, same files — but the context is gone. Before you can ask anything useful, the model has to re-read and re-familiarize. We watched a small team burn the first several minutes of every session on that overhead alone. Every day. Nobody was measuring it, so nobody was fixing it.

So we measured it — and built something to stop it. Semantic Cache Engine caches the familiarization context locally. Warm sessions answer in ~41ms, the cache gets smarter the longer you use it, and when a teammate imports your bundle, they start day one already warm.

Local-first, no server, no API key. VS Code first — everything else follows.

Local-first
Your cache lives on your machine. Nothing leaves without an explicit export. netstat-verified zero outbound connections during local operations.
Gets smarter over time
Every agent handoff commits a new card. Next time that question comes up — exact wording or a paraphrase — it's a cache hit. You only pay once.
Team multiplier
One developer's work warms the whole team. Export a bundle, share it once — every new hire skips the cold-start period entirely.

Your AI rebuilds its understanding of your codebase from scratch. Every session. Every developer.

📄
Session starts
Copilot pulls your most relevant files and chunks into context — hundreds to thousands of tokens — before it can answer anything.
🔄
LLM processes it
That context is gone the moment the session ends. Each developer re-establishes the same foundational understanding of the codebase independently, every day.
💰
You pay again
Different questions, same overhead. The re-familiarization cost hits every developer, every session, before a single line of useful work.
With Semantic Cache Engine
Session starts
Engine checks the local cache. Context is already there.
📌
Cache hit
Answer returned in ~41ms. No LLM call. Zero tokens used.
🆕
Team starts warm
75–78% of session-start tokens never leave your machine. Import a teammate's bundle — skip the cold start entirely.

How it works.

01

One slow session, then done

Install the extension. It indexes your codebase once and builds a local reasoning cache in .reason/. No account. No API key. That's the only cold session.

02

Every question hits the cache first

Before anything reaches the model, the resolver checks for a match — exact, paraphrase, or close-enough context. Most questions already have an answer. Hit: back in under a second, tokens: zero.

03

Not confident? It asks, then remembers

When confidence is too low, it routes to the full agent with a compacted context prompt. The answer comes back, a new card is committed. Next time that question comes up, it's a cache hit. The longer you use it, the less it needs to ask.

04

Ship your cache to the team

Export a scrubbed bundle. Teammates import it. They start their first session with weeks of your context already loaded — no cold start, no re-explaining the codebase from scratch.

Recall
Exact match or paraphrase — answered from cache in ~41ms. No model call, zero tokens.
Graph-assisted
The engine walks your dependency graph, pulls the minimum relevant context, and sends a compressed prompt. Fewer tokens in, same quality out.
Agent handoff
New territory — full agent, full context. The answer is committed to cache before the session ends. You only pay once.

Claims we stand behind.

75–78%
Session-start token reduction
Measured with real OpenAI API calls (gpt-4o-mini), verified on two separate machines. Engine-on vs engine-off A/B test.
Claim 1b — real API measurement
0
Third-party outbound connections
Verified via netstat on two machines during local operations. The extension includes a built-in no-phone-home check you can run yourself.
netstat-verified
Day 1
Warm start for new teammates
Export from one session, import to a fresh workspace: 10/10 cache hits on first run. Every new developer skips the cold-start period entirely — verified in the Docker demo, reproducible in under 5 minutes.
Docker demo — try it live
Local
All reasoning stays on your machine
The cache lives in .reason/ in your workspace. Bundles only leave when you explicitly export and share them.
Architecture guarantee
The savings multiply. The 75–78% figure is per developer, per session. Multiply by team size, multiply by sessions per day. When teams share bundles, every new person starts warm from day one — on a codebase they've never opened. And the cache keeps getting smarter: every agent handoff commits a new card, so tomorrow's hit rate is higher than today's.

Cold session → export → import → warm session.

docker pull ghcr.io/leangridlabs/semantic-cache-engine-demo:latest
docker run -it --rm ghcr.io/leangridlabs/semantic-cache-engine-demo

Walks through ingest → cold run → export → import → warm run interactively. No account required.

Be first in.

The free dev tier is in final prep. Leave your email — we'll send one message when it ships, share the baseline measurement methodology, and you can run your own before/after numbers on day one.

No spam. No data sharing. Unsubscribe any time.