Your Data Never Leaves
Compression happens locally. No API call to an outside model, no context sent to a third-party server. Your documents, memory, and conversation history stay on your hardware.
Focus everything an agent must read into the smallest bright signal that still carries the full meaning.
A lighthouse Fresnel lens takes a small lamp and focuses it into a beam powerful enough to reach ships ten miles out. Beam does the same thing for the context an AI agent reads: it compresses vast knowledge down to the bright, essential signal. Locally, on your hardware, so your data never leaves.
Learn moreEvery AI agent works inside a context window, a fixed amount of reading space. Fill it with sprawling documents, conversation history, and tool output and the agent either slows to a crawl, misses the important parts, or runs out of room entirely. Token costs and latency compound on top. The category-level goal is large token reduction without signal loss. That's what Beam is built to deliver.
The Fresnel lens metaphor is exact: concentric rings of glass, each angled to redirect light that would otherwise be wasted, all pointing toward the same focused outgoing beam. Beam applies the same principle to tokens, compressing wide scattered context into a narrow, far-reaching signal.
The compression engine at the heart of Beam runs on a local model, the same machine your data already lives on. If that local model is unreachable, Beam refuses the job rather than quietly routing your context to a cloud service. That's fail-closed: the safe default is always stop, never leak.
Compression happens locally. No API call to an outside model, no context sent to a third-party server. Your documents, memory, and conversation history stay on your hardware.
If the local model is unreachable, Beam returns an error. It does not silently fall back to a cloud path. You always know which route was taken.
Unreachable local model → hard stop, not a cloud detour.
Not all context is equally important. Beam layers what it delivers: facts most critical to the immediate task come first, background and supporting material compressed further or summarized. The full meaning still arrives; the noise does not.
The compressed context store is planned to live on the Lighthouse Vault, the same own-your-data infrastructure that backs all of our sovereign tooling. One place, one trust model, your hardware.
Most context-compression tools work as a cloud proxy: your data leaves your machine to be summarized, then a smaller version comes back. For many workloads that's fine. For work that is genuinely sensitive (client files, proprietary code, personal memory) the moment your data leaves is the moment you lose full control. Beam is built for those workloads.
Beam is in active development. Here is what exists today and what is still being designed.
The local-model compression path is built and tested inside Hermes. It compresses context locally, is fail-closed (refuses if the local model is unreachable), and has been exercised in real agent runs.
An early build now exists: a standalone tool with a command line and an MCP server, so any agent (Claude Code, Cursor, Hermes) can compress through Beam -- local-only, fail-closed, with a reversible cache. A transparent proxy, a managed context store on the Vault, and a purpose-trained compression model are still on the roadmap. Not yet a packaged release.
We are not accepting sign-ups or quoting pricing yet. If you are working on a project where sovereign context compression matters to you, reach out. We are interested in the right early conversations.
Focus the lamp. Reach further. Keep the data home.
A Fresnel lens doesn't make the lamp brighter. It wastes less of the light that was already there. Beam applies the same logic to AI context: your knowledge is already there, your agent just needs to see it clearly, without flooding the window or sending it over the wire.