Nema Language

Experimental agent-oriented language for compiling NeuroState into control flow.

"Agents don't just compute. They feel."

---

Browser Demo (WebAssembly)

Open browser_demo/index.html directly in any modern browser — no server needed.

The demo compiles two agents (Emilia and Kernel) from .nema source to WASM. Each slider writes directly to a WASM f64 global; gate conditions are re-evaluated in real time. Drag a field below its threshold and watch the gate flip from ✅ to ❌ instantly.

Emilia_explore()    requires dp > 0.6     → dp + ac
Emilia_rest()       requires s > 0.5 ∧ gaba > 0.4  → s + gaba
Emilia_connect()    requires ox > 0.6     → ox + s
Emilia_mood_score() (no gate)             → Σ all fields

To regenerate after editing demo.nema:

python3 nema.py browser_demo/demo.nema --wasm
python3 browser_demo/gen.py

---

What is Nema?

Nema is a research prototype language where every agent carries a NeuroState — a 6-dimensional affective state based on neurotransmitters (dopamine, serotonin, acetylcholine, oxytocin, GABA, endorphin). Emotional state gates function execution, propagates between agents, decays over time, and drives memory management.

@requires(dp > 0.6) compiles directly to fcmp ogt + conditional branch in LLVM IR — not a runtime flag, not a config value, machine code.

One of the first experimental languages to compile agent affective state (NeuroState) into executable control flow via LLVM IR.

---

Install

pip install nema-lang

Requires Python 3.10+ and LLVM (via llvmlite).

---

Quick Start

# Run the REPL interpreter
nema hello.nema

# JIT compile and run benchmarks
python -m jit_run hello.nema

# Generate LLVM IR
nema hello.nema --compile
# → produces hello.ll

# Compile to WebAssembly (WAT + WASM binary via wabt)
nema hello.nema --wasm
# → produces hello.wat + hello.wasm

---

Language Example

agent Neko {
  mood: NeuroState = {
    dp: 0.8, s: 0.5, ac: 0.7,
    ox: 0.6, gaba: 0.4, e: 0.6
  }

  // compiles to: fcmp ogt double %dp, 6.000000e-01
  @requires(dp > 0.6)
  fn explore(path) { }

  // multi-condition gate — both must be true
  @requires(gaba > 0.5 and s > 0.4)
  fn sleep() { }
}

agent Kernel {
  mood: NeuroState = {
    dp: 0.5, s: 0.7, ac: 0.9,
    ox: 0.4, gaba: 0.5, e: 0.5
  }

  // real malloc — gated by emotional focus
  @requires(ac > 0.8 and dp > 0.3)
  fn alloc(size: i64) -> ptr<i64> { }

  @requires(ac > 0.5)
  fn write(addr: ptr<i64>, val: i64) -> void { }

  @requires(dp > 0.3)
  fn read(addr: ptr<i64>) -> i64 { }

  fn free(addr: ptr<i64>) -> void { }
}

---

Architecture

Nema source (.nema)
       ↓
   Lexer / Parser
       ↓
   Type Checker  ← validates NeuroState fields, ranges, always-fail gates
       ↓
      AST
       ↙        ↘
 Interpreter    LLVM Compiler
 (REPL mode)   (JIT / .ll output)
       ↓              ↓
  Runtime         Machine code
(emotion lives)  (emotion compiled)

---

The 6 Dimensions

Symbol	Neurotransmitter	Meaning
dp	Dopamine	Curiosity, motivation
s	Serotonin	Stability, calm
ac	Acetylcholine	Focus, attention
ox	Oxytocin	Trust, empathy
gaba	GABA	Inhibition, composure
e	Endorphin	Joy, achievement

---

REPL Commands

Command	Description
show <agent>	Display NeuroState + memory
trace <agent> [n]	Show last n emotion-change events as a table (rich, default 5)
call <agent> <fn> [args]	Call function (emotion-gated)
attract <A> <B> [strength]	Set symmetric attraction between agents
remember <agent> <key> <value>	Write to working memory
recall <agent> <key>	Retrieve from memory
introspect <agent>	Verbalize emotional state in Japanese
empathize <A> <B>	A absorbs 30% of B's emotional state
log <agent> <msg>	Log with emotional context
rand_mood <agent>	Randomize NeuroState
summarize <agent>	Swap working memory to long-term storage
spawn <agent> <fn>	Run agent function in background thread
threads	Show all active threads
transfer <from> <to> <var>	Transfer ownership between agents
shii <agent>	Inject しーちゃん spirit.db → NeuroState

---

Core Features

Emotion Gates → Machine Code

@requires(dp > 0.6)
fn explore(path) { }

@requires(ac > 0.8 and dp > 0.3)
fn alloc(size: i64) -> ptr<i64> { }

Each condition compiles to fcmp ogt + and i1 + conditional branch in LLVM IR.
Gate-rejected functions return -1 (or null for pointer types).

Real Memory Operations

alloc, write, read, free compile to actual malloc/store/load/free instructions — not simulated, real machine code guarded by emotional state.

Static Type System

fn alloc(size: i64) -> ptr<i64> { }
fn write(addr: ptr<i64>, val: i64) -> void { }

Supported types: i64, i32, f64, bool, void, ptr<T>, NeuroState

Static Type Checker

Validates at parse time:

• Unknown NeuroState fields → error
• Values outside [0.0, 1.0] → error
• @requires(dp > 1.5) → warning (always fails)

Emotional Decay (background thread)

Every 5 seconds, all emotions decay at neurotransmitter-specific rates. Serotonin fades slowly; dopamine faster. Agents grow tired if left alone.

Agent Attraction

attract Neko Shii 0.5

Symmetric emotional pull — agents converge toward each other's state on every tick.
Compiles to (B[f] - A[f]) * strength * 0.1 delta applied symmetrically via LLVM fsub/fmul/fadd.

CPOS Memory Layer

Working memory (RAM, max 5 entries) + long-term storage (JSON).
When gaba ≥ 0.7, composure triggers automatic memory consolidation (swap to disk).

For Loops

for i in 0..5 { log(i) }             // range loop
for val in [1.0, 2.0, 3.0] { log(val) } // list loop

@requires(dp > 0.5)
fn gated_loop() -> f64 {
  for i in 0..3 { log(i) }  // gate rejects entire loop if dp ≤ 0.5
  return dp
}

Result<T> Error Type

fn divide(a: f64, b: f64) -> f64 {
  branch b == 0.0 {
    let r = err("divide by zero")
    log(r)
    return 0.0
  } else {
    let r = ok(a)
    log(r)
    return a
  }
}

fn try_it() -> f64 {
  let result = ok(42.0)
  match result {
    ok(v)    { return v }
    err(msg) { log(msg); return 0.0 }
  }
  return 0.0
}

Typed Channels

let ch = channel<i64>
spawn worker(ch)
send ch 42
recv ch -> val { log(val) }
close ch

set<T> Type (v0.6.0)

let tags = set<i64>
tags.add(42)
tags.add(42)          // duplicates silently ignored
print(tags.size())    // 1
print(tags.contains(42))  // True
tags.remove(42)
let items = tags.to_list()

Supported methods: .add(v), .remove(v), .contains(v) -> bool, .size() -> i64, .to_list() -> list<T>

map<K,V> Type (v0.7.0)

let scores = map<i64, i64>
scores.set(10, 100)
scores.set(20, 200)
print(scores.get(10))       // 100
print(scores.contains(20))  // True
scores.remove(20)
print(scores.keys())        // [10]
print(scores.values())      // [100]

Supported methods: .set(k, v), .get(k), .remove(k), .contains(k) -> bool, .size() -> i64, .keys() -> list<K>, .values() -> list<V>

Emotion Trace Visualization (v0.7.0)

> trace Neko
              Neko
┏━━━━━━━━━━━━━━━━┳━━━━━━━┳━━━━━━━┓
┃ 発生源          ┃ field ┃     Δ ┃
┡━━━━━━━━━━━━━━━━╇━━━━━━━╇━━━━━━━┩
│ @after:explore │ dp    │  +0.1 │
│ @after:explore │ e     │ +0.05 │
└────────────────┴───────┴───────┘

Renders the same change history as mood.origin(n) (used inside .nema code) as a rich table in the REPL, color-coded by sign (green=increase, red=decrease). Falls back to the plain-text format if rich isn't installed.

Standard Library (v0.6.0)

import "lib/math.nema"
import "lib/collections.nema"

agent Main {
  fn run() -> void {
    let m = Math()
    print(m.abs(0 - 5))         // 5
    print(m.pow(2, 8))          // 256
    print(m.clamp(15, 0, 10))   // 10

    let c = Collections()
    let nums = [1, 2, 3, 2, 1]
    c.unique(nums)              // unique count: 4
  }
}

lib/math.nema: abs, max, min, clamp, pow
lib/collections.nema: unique, intersection_size
lib/io.nema: print_list, print_hr, print_labeled
lib/sensors.nema: sample (capability-gated read, demonstrates cross-agent import)

Contracts (Layer 9)

agent Counter {
  contract { dp >= 0.1  gaba >= 0.1 }  // checked after every call/tick
}

Violations raise a ContractError and are logged with the offending field.

JIT Performance

Operation	Speedup over interpreter
tick (decay)	10.4×
gate check	6.3×
attract	9.6×

---

LLVM IR Output

; NeuroState as double[6]
@"mood_Neko" = internal global [6 x double] [
  double 0x3fe999999999999a,  ; dp = 0.8
  double 0x3fe0000000000000,  ; s  = 0.5
  ...
]

; @requires(dp > 0.6) → fcmp ogt
define i32 @"fn_Neko_explore"() {
entry:
  %val_dp = load double, double* %dp_ptr
  %cmp_dp = fcmp ogt double %val_dp, 6.000000e-01
  br i1 %gate, label %exec, label %reject
exec:
  ret i32 0
reject:
  ret i32 -1
}

; alloc: real malloc gated by emotion
define i64* @"impl_Kernel_alloc"(i64 %size) {
entry:
  %cmp_ac = fcmp ogt double %ac, 8.000000e-01
  %cmp_dp = fcmp ogt double %dp, 3.000000e-01
  %gate = and i1 %cmp_ac, %cmp_dp
  br i1 %gate, label %exec, label %reject
exec:
  %nbytes = mul i64 %size, 8
  %raw = call i8* @malloc(i64 %nbytes)
  %ptr = bitcast i8* %raw to i64*
  ret i64* %ptr
reject:
  ret i64* null
}

---

Static Type Checking

nema myfile.nema --check

Catches errors at parse time, before any execution:

[WARN]  Neko.explore: @requires(dp > 1.5) — always fails (dp max is 1.0)
[ERROR] Kernel.alloc: unknown NeuroState field 'motivation' (use: dp s ac ox gaba e)
[ERROR] Kernel.alloc: NeuroState value 1.8 out of range [0.0, 1.0]

Exit code 0 = clean, 1 = errors found.

---

Safety Model

Nema has six layers of execution safety. No single layer is sufficient — they compose.

Layer 1:  Emotion Gate       @requires(dp > 0.6)  → fcmp ogt in LLVM IR
Layer 2:  Post-condition     @ensures(gaba > 0.3) → verified after execution; runs @on_error on fail
Layer 3:  Fallback           @on_error { ... }    → runs on gate fail OR ensures fail
Layer 4:  Static Type Check  unknown fields / out-of-range values → compile-time error
Layer 5:  Ownership          own / release / recv — double-free raises serotonin penalty
Layer 6:  Capability         capability: { alloc, emit } — privileged ops (alloc/free) require declaration
Layer 7:  Trust Score        trust: { AgentB: 0.8 } — query/send blocked if trust < 0.3
Layer 8:  Memory Isolation   CPOS working memory (max 5) / long-term JSON / auto-swap
Layer 9:  Contract           contract { dp >= 0.1 } — invariant checked after every call/tick
Layer 10: CPOS Gate          @cpos_gate — blocks on NeuroState WARNING or suspicious call pattern

Emotion gates express agent readiness, not permissions. Capabilities enforce permissions. Trust enforces identity. All three compose.

CPOS Gate (Layer 10)

agent SecureAgent {
  mood: NeuroState = { dp: 0.5, s: 0.6, ac: 0.5, ox: 0.5, gaba: 0.3, e: 0.5 }

  @cpos_gate
  fn export_data() {
    let src = mood.origin()   // trace what caused the current mood state
    log(src)
    emit "data exported"
  }
}

@cpos_gate triggers when either condition is met:

• gaba >= 0.6 — NeuroState WARNING level (accumulated context poisoning)
• Alternating high/low impact call pattern detected over the last 6 calls (S6-class attack)

mood.origin() returns the trace of what changed the mood state and from where:

[mood.origin]
  ← user_input:hello: gaba+0.15, e+0.1
  ← fn:do_work: s-0.1
  ← user_input:suspicious: gaba+0.12, e+0.05

This closes the gap that scalar gates (C4) and trajectory gates (C5) both leave open.

agent Kernel {
  capability: { alloc, free, write, read, emit }
  trust: { Process: 0.8 }

  @requires(ac > 0.8)      // Layer 1: must be focused
  @ensures(gaba > 0.3)     // Layer 2: must remain calm after
  @on_error { emit kernel_fail 1 }  // Layer 3: fallback if either fails
  fn alloc_buf(size: i64) -> ptr<i64> {
    own buf = alloc(size)  // Layer 5+6: owned + kernel-only
    return buf
  }
}

---

Examples

File	Demonstrates
hello.nema	Emotion gates, when blocks, agent attraction
memory.nema	CPOS working / long-term memory, gaba-triggered swap
kernel.nema	Emotion-gated malloc / write / read / free, ownership transfer
concurrent.nema	Multi-agent concurrency, mailbox recv, spawn
for_demo.nema	for loops (range & list), emotion-gated iteration
result_demo.nema	Result<T> error type, ok(v) / err(msg), match
channel_demo.nema	Typed channel<T>, send / recv / close across agents
contract_demo.nema	Layer 9 contracts — NeuroState invariants
set_demo.nema	set<T> type — .add/.remove/.contains/.size/.to_list
map_demo.nema	map<K,V> type — .set/.get/.remove/.contains/.size/.keys/.values
stdlib_demo.nema	Standard library — Math, Collections
match_demo.nema	match expressions, multi-file import
typed.nema	Typed function signatures, ptr<T>

---

Testing

pip install -e ".[dev]"
pytest

tests/test_demos.py type-checks and runs every .nema file under repo root against regressions (no crashes, no --check errors), plus targeted assertions for set<T>, map<K,V>, and capability-gated alloc.

---

Files

File	Role
lexer.py	Tokenizer
parser.py	AST parser
ast_nodes.py	AST node definitions
typechecker.py	Static type checker
evaluator.py	Interpreter runtime + concurrent execution
stdlib.py	Standard library (introspect, empathize, CPOS)
compiler.py	LLVM IR code generator
jit_run.py	JIT compiler + runner
nema.py	Entry point + REPL
benchmark.py	JIT vs interpreter benchmarks
shiichan.py	しーちゃん spirit.db → NeuroState bridge

---

Background

Nema is built on two original concepts:

• NeuroState — A 6-dimensional emotional model based on neurotransmitters, developed as part of the Emilia OS research project. [Zenodo DOI: 10.5281/zenodo.19734147]
• CPOS (Context Pointer OS) — A cognitive memory kernel for LLM agents. [github.com/kagioneko/context-pointer-os]

---

Nema — where code has feelings.