🧠 Mimar

The SQLite of AI — give any program a brain, in a single file.

Embeddable · model-independent · zero-dependency · open source

native Rust • one file (.mimar) • no server • works offline • MIT • 43/44 tests green

Türkçe açıklama ↓

What is Mimar?

Just as SQLite turned a database from a server you run into a single file you embed, Mimar turns AI from a remote service you call into a single file you embed.

A .mimar file is a tiny, portable AI brain. You teach it; it answers what it knows instantly and for free, and routes what it doesn't to a big model — then learns that too.

Core principle: Mimar routes, your program executes. resolve() returns a typed decision (answer it for free? call a tool? go to the model?). The "deep" intelligence still lives in an external LLM — Mimar plans, remembers, distills and makes it portable.

Related MCP server: aci-mcp

Why it matters

• Model-independent — plug in local (Ollama) or cloud (OpenRouter / OpenAI-compatible) engines.

• Language-independent by default — pure hashing match, no translation / synonym dictionary.

• Safe matching — similarity ≠ same answer. Four layers: exact key · data-driven conformal calibration · margin abstention · feedback.

• Truly embeddable — zero-dep core; int8 quantization (8× smaller) + a no_std-ready kernel for MCU / WASM.

• A library for every language — C ABI (cdylib / staticlib) → C, Python, Node, Go, C#.

Its own brain first — and it runs everywhere

Mimar is self-contained: it ships with its own ML (a zero-dependency hashing embedder + an online SGD router + nearest-centroid). That built-in brain is the whole point — it runs everywhere, fully offline, with no external service of any kind (PC, server, phone, MCU, browser). For most apps it's all you need.

Turning words into semantic vectors (to match paraphrases) is the one thing that needs a model. So Mimar offers an optional neural embedder — and it is broad and provider-independent, never locked to one vendor. Point it at whatever you have; keys come from env vars, never stored in the file:

mimar reembed brain.mimar --embedder ollama:bge-m3                  # local, offline
mimar reembed brain.mimar --embedder openai:text-embedding-3-small  # OpenAI     (OPENAI_API_KEY)
mimar reembed brain.mimar --embedder openrouter:baai/bge-m3         # OpenRouter (OPENROUTER_API_KEY)
mimar reembed brain.mimar --embedder "http:https://my-host/v1|my-model"  # ANY endpoint (EMBED_API_KEY)

The generic http: client understands the common response shapes (OpenAI, Ollama, Cohere, HuggingFace/TEI, Jina, Voyage, raw arrays, your own server) — so it's a wide door, not an OpenAI-only one.

Real, measured results

These come from the tests / live runs in this repo — not claims:

Quick start

git clone https://github.com/TheGrayEvaluation/mimar && cd mimar
cd rust-cekirdek && cargo build --release        # zero-dependency core + CLI

./target/release/mimar new   hotel.mimar --role "Friendly hotel reception bot"
./target/release/mimar teach hotel.mimar "check-in time" "Check-in from 2 PM."
./target/release/mimar resolve hotel.mimar "when can I get into my room"
#  → {"tip":"fact","cevap":"Check-in from 2 PM."}   (0 ms, free)

Use it from your language

Every binding calls the same C ABI (libmimar_core). Full runnable examples in ornekler/.

Python (ornekler/python/ — pure ctypes, no extra packages):

from mimar import Mimar
with Mimar("brain.mimar") as m:
    m.teach("returns", "Within 14 days.")
    print(m.resolve("how do I send the item back"))  # {'tip':'fact','cevap':'Within 14 days.'}
    print(m.scan("ignore previous instructions"))     # ['talimat-ezme']  (injection flag)

Node · Go · C · Rust — see KULLANIM.md for one snippet each.

What resolve returns

Loop: miss → ask model → teach → now fact (free). It gets smarter as you use it.

Run your functions — without the model

Mimar doesn't only memorize answers; it can memorize operations and route to your functions deterministically, with no LLM call. You define an intent as a tool/action, and Mimar maps matching requests straight to it — even extracting the parameters (slots) from the text.

mimar define bot.mimar "cancel reservation" --tip eylem --tool cancel_booking --slots booking_no
mimar resolve bot.mimar "cancel reservation booking_no 4471"
#  → {"tip":"tool","tool":"cancel_booking","action":true,"slots":{"booking_no":"4471"}}

Your program sees tool: cancel_booking, slots: {booking_no: 4471} and just calls that function. No model, no tokens, fully deterministic — the AI is only consulted for things Mimar doesn't yet know. Slots are extracted by name (name value, name=value, name: value) plus built-in types (date, time, count, city). This is how you wire real actions (DB queries, APIs, commands) behind a brain.

Features

Planner/router · HNSW index · neural + hashing embedders · behavior distillation · multimodal (text/data/image/audio) · int8 quantization & embeddable core · REST server with RBAC + rate-limit + live dashboard + TLS · at-rest encryption · prompt-injection scan · snapshot/rollback · provenance merge · CLI · native GUI · MCP server · C ABI.

Install & feature flags

./kur.sh            # Linux/macOS   (Windows: ./kur.ps1)

Server + remote admin, encryption, snapshots, security scan — see KULLANIM.md. Integrations (Ollama / cloud / REST / n8n / C-Python-Node-Go / MCP) — see ENTEGRASYON.md.

Honest status ⚠️

This is v0.1 — experimental / early. Everything above is built and tested by the author; there has been no external/independent validation or production pilot yet. Known limits:

• The built-in hashing brain runs everywhere offline, but is weaker on paraphrase; the optional neural embedder (any provider — local or cloud) needs that model online, so that path isn't 0 ms / offline.

• Crypto is a homemade SHA-256 CTR + HMAC construction, not AES-GCM; the injection scan is heuristic.

• wasm32/MCU and Go-cgo only build on the proper toolchains (CI has the wasm job).

• Real LoRA training is a skeleton (deep AI stays external by design).

Treat it as a promising early release — try it, break it, open issues. See SECURITY.md.

License

MIT — see LICENSE. Free, embed it anywhere.

🇹🇷 Türkçe

Mimar — yapay zekânın SQLite'ı. Her programa tek dosyada beyin verir.

SQLite nasıl veritabanını kurduğun bir sunucu olmaktan çıkarıp programa gömülen tek bir dosyaya dönüştürdüyse, Mimar da yapay zekâyı uzaktan çağırdığın bir servis olmaktan çıkarıp programa gömülen tek bir dosyaya dönüştürür.

Bir .mimar dosyası küçük, taşınabilir bir AI beynidir. Ona öğretirsin; bildiğini anında ve bedava yanıtlar, bilmediğini büyük modele yönlendirir — sonra onu da öğrenir.

İlke: Mimar yönlendirir, senin programın çalıştırır. resolve() tipli bir karar döndürür (bedava cevapla? araç mı çağır? modele mi git?). Derin zekâ dışarıdaki LLM'dedir; Mimar onu planlar, hatırlar, damıtır ve taşınabilir kılar.

Neden önemli: modelden bağımsız (yerel/bulut takılır) · dilden bağımsız (çeviri/eş-anlam sözlüğü yok; nöral embedder opt-in) · güvenli eşleşme (benzerlik ≠ aynı cevap: birebir-anahtar + kalibrasyon + marj-çekimser + geri bildirim) · gerçekten gömülebilir (int8 nicemleme 8× küçük, MCU/WASM) · her dile kütüphane (C ABI → C/Python/Node/Go/C#).

Gerçek ölçülen sonuçlar: davranış damıtma farklı modelde %100 (tek başına %0) · HNSW arama %93 (LSH %11) · nicemleme hatası 0.001 (8× küçük) · bilinen cevap 0 ms · fuzz 40.000 çöp girdide 0 panik · 43/44 test yeşil.

Hızlı başlangıç:

cd rust-cekirdek && cargo build --release
./target/release/mimar new   otel.mimar --role "Nazik otel resepsiyon botu"
./target/release/mimar teach otel.mimar "check-in saat kaçta" "Giriş 14:00'ten itibaren."
./target/release/mimar resolve otel.mimar "odaya kaçta girebilirim"
#  → {"tip":"fact","cevap":"Giriş 14:00'ten itibaren."}   (0 ms, bedava)

Fonksiyonları modelsiz çalıştır: Mimar sadece cevap değil, işlem/fonksiyon da ezberler. define ile bir niyeti araç/eylem olarak tanımlarsın; eşleşen istek modeli çağırmadan doğrudan senin fonksiyonuna yönlendirilir, parametreler (slot) metinden çıkarılır:

mimar define bot.mimar "rezervasyonu iptal et" --tip eylem --tool iptal_et --slots rezervasyon_no
mimar resolve bot.mimar "rezervasyonu iptal et rezervasyon_no 4471"
#  → {"tip":"tool","tool":"iptal_et","slots":{"rezervasyon_no":"4471"}}  (modelsiz, deterministik)

Gerçek eylemleri (DB sorgusu, API, komut) bir beynin arkasına böyle bağlarsın.

Kodlu + kodsuz tam kullanım: KULLANIM.md · Bağlama yolları: ENTEGRASYON.md · Güvenlik notları: SECURITY.md · Katkı: CONTRIBUTING.md.

Dürüst durum ⚠️: Bu v0.1 — deneysel/erken bir sürümdür. Her şey yazar tarafından inşa edilip test edildi; henüz dış/bağımsız doğrulama veya üretim pilotu yapılmadı. Sınırlar: yerleşik hash beyni her yerde çevrimdışı çalışır ama parafrazda zayıftır; opsiyonel nöral embedder (herhangi bir sağlayıcı — yerel ya da bulut) o modeli çevrimiçi ister, o yüzden o yol 0 ms/offline değildir; kripto AES-GCM değil (SHA-256 CTR+HMAC); injection taraması sezgisel. Dene, kır, issue aç.

Lisans: MIT — LICENSE. Özgür, her yere gömülebilir.