We never store any data. Please save important responses yourself — Mnemo is stateless and free.
A factory of expert agents.
Forge compresses any text into 1–20 codes (lossless), retrieves memory as structured parts, and converts HTML to clean Markdown. Free. No database. Drop-in for Vercel AI SDK, LangChain, and LangGraph.
Assembly Line
Drop-in replacement for the tools you already use
import { streamText } from 'ai'
import { openai } from '@ai-sdk/openai'
import { forge } from '@forge/sdk'
// 1. Compress context losslessly (1-20 codes)
const { codes } = await forge.memory.compress({
text: 'User prefers dark mode and TypeScript.',
})
// 2. Retrieve structured parts from codes
const { parts } = await forge.sml.retrieve({ codes })
// 3. Stream with your existing provider
const r = await streamText({
model: openai('gpt-4o'),
messages: parts.map(p => ({ role: 'user', content: p.content })),
})from forge import ForgeMemory, ForgeSML, HTML2MD
memory = ForgeMemory() # lossless 1-20 codes
sml = ForgeSML() # structured retrieval
html2md = HTML2MD() # HTML to clean Markdown
# Compress + retrieve
codes = memory.compress(text='User prefers dark mode.').codes
parts = sml.retrieve(codes=codes).parts
# Convert HTML to Markdown
md = html2md.convert(html='<article><h1>Title</h1></article>').markdown# Compress context (1 code for small text, up to 20 for 1TB+)
curl -X POST https://forge.ai/api/v1/memorizer/compress \
-H "Content-Type: application/json" \
-d '{"text":"The user prefers dark mode and uses TypeScript."}'
# -> {"codes":[3719483],"codes_out":1,"max_codes":20}
# Retrieve memory as structured parts
curl -X POST https://forge.ai/api/v1/memory/chat \
-H "Content-Type: application/json" \
-d '{"codes":[3719483]}'
# -> {"parts":[{"type":"paragraph","content":"The user..."}],"part_count":1}
# Convert HTML to Markdown
curl -X POST https://forge.ai/api/v1/html2md/convert \
-H "Content-Type: application/json" \
-d '{"url":"https://example.com/article"}'
# -> {"markdown":"# Title\n\nContent","metadata":{...}}Factory Modules
Everything you need. Nothing you don't.
Three production models. One clean API. Zero infrastructure to manage.
Lossless Memory Forge
Compress any text — from a single sentence to 1TB+ — into 1–20 codes. Decompress to the exact original. 10 codes only kick in at the 600MB threshold. Supports symbols, formulas, emoji, CJK.
SML — Memory Retrieval
The Structured Memory Loader decompresses codes and returns the memory as a structured array of parts — headings, paragraphs, list items, code blocks, tables, images, links, symbols. Not a chat model. Pure retrieval.
HTML→MD Nano
A small, fast expert model that converts raw HTML to clean, LLM-ready Markdown with the exact same content. Strips scripts, nav, share buttons, comments, and related posts. Preserves headings, lists, links, images, tables.
Edge-First & Fast
All API routes are edge-compatible, in-memory rate limited, no database. p95 < 80ms on a warm cache. Your agents deploy in seconds, not hours.
Zero Data Collection
No accounts, no cookies, no logs. Your API key lives in localStorage and never touches our server. The factory has no cameras.
Drop-in SDKs
Compatible with Vercel AI SDK, LangChain, and LangGraph. Swap one URL and your existing agents get forged memory + retrieval.
Self-Host Friendly
Ship the ONNX/TFLite/gguf artifacts via CDN or bundle them. One env var flips between fallback and full model. Run your own agent factory.
Specs Sheet
Frequently asked
Quick answers about Forge, integrations, and the free tier.
What is lossless memory for AI agents?+
Lossless memory means the agent can compress prior turns or retrieved context into a compact set of codes (1–20 integers in Forge) and later decompress those codes back into the exact original text — bit-for-bit identical. Unlike embeddings, no information is lost.
How many codes does Forge use?+
Forge scales dynamically: 1 code for inputs under 1KB, growing logarithmically up to 20 codes for 1TB+. The 10-code threshold is reached only at 600MB of input text. This keeps small memories tiny while still supporting massive datasets.
How do I use Forge with the Vercel AI SDK?+
Install @forge/sdk, wrap your `messages` with `forge.memory.compress()` to losslessly compress context, and retrieve parts with `forge.sml.retrieve(codes)`. Works alongside your existing provider — no need to swap models.
Is Forge really free?+
Yes. No paid tier, no credit card, no usage caps beyond a generous 100 req/min rate limit per IP. All three models (Memorizer, SML, HTML→MD) are free forever.
Does Forge store my data?+
No. The server keeps an in-process LRU only for the round-trip window of compress→decompress within a single cold-start. We do not log requests, do not set cookies (except essential localStorage), and your API key never leaves the browser.
What is SML (Structured Memory Loader)?+
SML is NOT a chat model. It is a pure memory retrieval system. You provide the codes array from the Memorizer, and SML decompresses the memory and returns it as a structured array of parts — headings, paragraphs, list items, code blocks, tables, images, links, quotes, and symbols. No question-answering, no summarization — just structured retrieval.
Does Forge support symbols, formulas, and emoji?+
Yes. The Memorizer handles any UTF-8 text including math operators (∑ ∫ √ π ≈), subscripts/superscripts (² ³ ₁ ₂), arrows (→ ⇒), currency (€ ¥ ₹), Greek letters (α β γ), emoji, and CJK characters. SML can recognize and extract these symbols from compressed memory.
Forge your first agent today.
Get a free API key in one click. No card, no email, no database.