The AI engine that remembers.

Farnah.ai is a pre-release, locally runnable, open-weights native AI engine and agentic CLI. Persistent Spectrum Multi-Layer Memory distills, links, and recalls durable project knowledge across sessions, while OpenAI-compatible serving lets familiar clients connect to the same local engine. Bounded agent tools and MCP, multimodal components, and an eval-gated training path are composed around pluggable local models; accepted evidence can produce Farnah-native adapters and checkpoints without making any one bootstrap model the product. Local-first operation keeps control of models, memory, and data flow with the operator. The name comes from the Old Persian form farnah-: glory, splendor, and fortune associated with a luminous or fiery power. The Remembering Sun mark carries that idea into layers of memory.

Persistent memorySpectrum Multi-Layer Memory distills and recalls project knowledge across sessions, beyond the current context window.
Local open weightsThe pre-release stack runs qualified open weights through locally managed providers, with no required cloud model endpoint.
OpenAI-compatibleExisting compatible clients can connect to local serving. The farnah CLI adds bounded tools and MCP.

Talk to us about Farnah.ai →

Your mail goes to the team building the engine. You get a walkthrough and the current build status.

How it’s built

A composed engine, not one giant model: CLI and OpenAI-compatible serving connect bounded agents, Spectrum memory, task routing, local providers, multimodal components, and gated native artifacts.

Spectrum memory

The differentiator. Spectrum Multi-Layer Memory is a persistent subsystem with summary-first recall, deep_load on demand, topic links, and distilled write-back. It keeps model runtimes stateless while durable project knowledge survives session boundaries. Its facade stays stable as storage evolves, with bounded retrieval instead of global scans on the hot path.

Model router

Task-aware orchestration classifies code, conversation, reasoning, and vision requests, then selects a qualified local runtime and model behind one interface. Bootstrap models stay pluggable and revision-pinned; compatibility and held-out evaluations, not a fixed name list, decide what may be used.

Local serving & native artifacts

A provider boundary keeps local CPU and GPU runtimes replaceable while the OpenAI-compatible gateway stays stable. Farnah begins with qualified open weights, then uses redacted traces, Spectrum memory, and eval labels to build deterministic datasets. Adapters or checkpoints become Farnah-native artifacts only after training, held-out evaluation, and export gates pass.

Agentic CLI & MCP

A bounded tool loop — shell, file read/write, web fetch, sandboxed code exec — with MCP support for external tools, behind an OpenAI-compatible gateway. The farnah command provides REPL and one-shot modes; existing compatible clients talk to the same serving API.

Multimodal input

Pre-release multimodal components combine a frozen vision encoder with a trainable projector so image context can enter the engine without coupling the product to a fixed model. Multimodal artifacts advance through the same evaluation gates as the rest of the stack.

Why it exists

Own your engine. Keep your context.

No amnesiaContinuity across sessions via Spectrum memory, instead of re-explaining your project every time you open a new window.
No lock-inOpen weights, pluggable local providers, and a stable compatible API. There is no required cloud model endpoint, and optional external tools remain an operator choice.
Measured, not hypedRoutes, recalls, and tool actions create evidence. Memory policies, model choices, and native adapters or checkpoints advance only when evaluation gates demonstrate improvement.
Pre-releaseFarnah runs locally as an engineering system, but it is not presented as a public release. Integration, packaging, safety, and full-system evidence remain gated while the team works toward release.

References

Model Context Protocol (MCP)The open tool protocol Farnah bridges so external tools can plug into its bounded agent loop.
Ollama — local model servingA CPU-friendly local runtime supported behind Farnah’s provider boundary.
vLLM — high-throughput GPU servingA GPU serving path supported behind the same provider interface.
Encyclopaedia Iranica — farnah-Historical context for the Old Persian name element and its associations with glory, fortune, and luminous power.

Next case: Supply Chain Guard →