Memelang v11
Watch video on YouTube
Memelang is an AI-optimized query language that significantly reduces token count and model size for LLM text-to-SQL. The code below is designed to be copy-and-pasted into your LLM.
arXiv Paper · GitHub Repo · Patent Spec
document.getElementById('copy').style.background='',500);return false">Copy all code<br># info@memelang.net | (c)2026 HOLTWORK LLC | Patented<br># MEMELANG is a terse query DSL IR for LLM text-to-SQL<br># Axial grammar: Axis2 -> Axis1 -> Axis0 -> Cell<br># Whitespaces are syntactic and trigger "new Cell"<br># Never space between operator/comparator/comma/flag and values
MEMELANG_VER = 11.04
basic_syntax = '[table WS] [column WS] [":$" var][":" ("min"|"max"|"cnt"|"sum"|"avg"|"last"|"grp")] [":" ("asc"|"des")] ["" "\"" string "\""] [("="|"!="|">"|"="|"0; rating :DESC="Decimal 0-5 star rating of performance";:dec>0.0;0; name :DESC="Actor's full name";:str; age :DESC="Actor's age in years";:int>=0;0; description :DESC="Brief description of movie plot";:str; year :DESC="Year of production AD";:int>1800;=41; _;;
""" Role 567 and 8901 """<br>roles id 567,8901; _;;
""" Films with dystopian society narratives sim>.33 """<br>movies description "dystopian"=20;=4.2; actor :grp;;
""" Minimum role rating by actor, low to high """<br>roles rating :min:asc; actor :grp;;
""" Roles in movies mentioning robot rated 3+ """<br>movies description "robot"=3;;
""" Costars seen with Bruce Willis or Uma Thurman """<br>roles actor :$a~"Bruce Willis","Uma Thurman"; movie _;@ @ @; actor !$a;;
""" War stories before 1980: top 12 movies by minimum role rating """<br>movies year "war""robot"; year >=1900; "robot"; %col=year; >=1900; "robot" >=1900; "robot"; #year; >=1900; "robot"; :#year>=1900; ||'),<br>('CMP', r'>=||{p})" for k, p in CELL_PATTERN))
PAD_MODES = {'qry','tab'}<br>FLAG_KINDS = {'FLAG','BIND','EVAR','ASSN'}<br>LIT_KINDS = {'TIM','DEC','INT','ALN','QUO','EMB'}<br>VAR_KINDS = {'VAR','WLD','REL','EVAR','SLOT'}<br>DAT_KINDS = LIT_KINDS | VAR_KINDS<br>RELCOORD = {<br>'@0': ['-1','-1'],<br>'@1': ['-1','-2'],<br>'@2': ['-1','-3'],<br>'@3': ['-1','-4'],<br>'@4': ['-1','-5'],<br>'@' : ['-1','+0'],<br>'^' : ['-1','end','+0'],
# Atomic token<br>class Tok:<br>def __init__(self, kind: str, src: str, canon: Optional[str] = None):<br>self.kind = kind<br>self.src = src<br>canon = src if canon is None else canon<br>self.canon = CANON.get(canon) or canon<br>parser = {'QUO': json.loads, 'EMB': json.loads, 'DEC': float, 'INT': int}.get(kind)<br>self.dat = parser(src) if parser else src<br>def __str__(self): return self.src<br>def __repr__(self): return self.canon<br>def __eq__(self, other): return repr(self) == repr(other)<br>def __hash__(self): return hash(self.src)<br>def __bool__(self): return bool(self.src)
TOK_NULL = Tok('NULL', '')
# Sequence of tokens<br>class Seq(list[Tok]):<br>opr: Tok = TOK_NULL<br>def __init__(self, *items):<br>super().__init__(items)<br>self.opr = TOK_NULL<br>def __str__(self): return self.opr.src.join([str(t) for t in self if len(str(t)) or t.kind=='HOLD'])<br>def __repr__(self): return self.opr.src.join([repr(t) for t in self])
# Predicate expression<br>class Cell:<br>flag: Seq<br>left: Seq<br>comp: Tok<br>right: Seq<br>padded = False
def __init__(self, src: str):<br>self.left = Seq()<br>self.flag = Seq()<br>self.comp = Tok('EQL', '', '=')<br>self.right = Seq(Tok('WLD', '', '_'))
toks = []<br>for m in CELL_REGEX.finditer(src):<br>kind = m.lastgroup<br>text = m.group()<br>if kind == 'WS': continue<br>if kind == 'MISMATCH': raise Err(f'E_TOK {text!r}')<br>toks.append(Tok(kind, text))
i, n = 0, len(toks)
def peek(): return toks[i].kind if i = n: raise Err('E_EOF')<br>t = toks[i]<br>i += 1<br>return t
# FLAGS<br>while peek() in FLAG_KINDS:<br>self.flag.append(take())
# LEFT (prefix MOD)<br>if peek() == 'MOD':<br>self.left.opr = take()<br>self.left.append(Tok('HOLD', ''))<br>t = take()<br>if not t.kind in DAT_KINDS: raise Err('E_TERM_DAT')<br>self.left.append(t)
# COMPARATOR<br>if peek() == 'CMP':<br>self.comp = take()<br>if not peek() in DAT_KINDS: raise Err('E_DAT')
# RIGHT (values, OR-joined)<br>if peek() in DAT_KINDS:<br>self.right.clear()<br>while peek() in DAT_KINDS:<br>self.right.append(take())<br>if peek() == 'OR':<br>self.right.opr = take()<br>if not peek() in DAT_KINDS: raise Err('E_OR_TRAIL')
if i != n: raise Err(f'E_EXPR_TRAIL {toks[i:]}')
# PLACEHOLDER: OVERWRITE WITH YOUR EMBEDDING FUNCTION<br>def vectorize(self, tok: Tok) -> Tok:<br>if tok.kind == 'EMB': return tok<br>if tok.kind not in {'QUO', 'ALN'}: raise Err('E_EMBED')<br>return Tok('EMB', json.dumps([0.1, 0.2]))
@property<br>def single(self) -> Tok:<br>return self.right[0] if self.comp.canon == '=' and len(self.right) == 1 else TOK_NULL
@property<br>def literal(self) -> Tok:<br>tok = self.single<br>return tok if tok.kind in LIT_KINDS else TOK_NULL
def find(self, kind:str) -> Tok:<br>return next((flag for flag in self.flag if flag.kind == kind), TOK_NULL)
def bind(self, tok: Tok):<br>if tok not in self.flag: self.flag.append(tok)
def __str__(self) -> str: return f"{self.flag}{self.left}{self.comp}{self.right}"
def __repr__(self) -> str: return f"{self.flag!r}{self.left!r}{self.comp!r}{self.right!r}"
def __bool__(self)...